SO_LINGER brokenness...
Sean Chittenden
sean at chittenden.org
Wed Jan 23 12:20:36 PST 2008
Howdy. Looks like SO_LINGER is broken on loopback. Here's the test
case:
0. Install memcached (databases/memcached)
1. Download test_linger.c: fetch http://sean.chittenden.org/pubfiles/freebsd/test_linger.c
2. Compile: gcc test_linger.c -o test_linger
3. Fire up memcached in a new window: memcached -vvv -p 11211
4. Test: ./test_linger localhost 11211
5. Watch for either:
<16 new client connection
<16 set valid_key 0 0 8
>16 STORED
<16 set valid_key 0 0 8
>16 STORED
<16 connection closed.
Fail.
or
<16 new client connection
<16 set valid_key 0 0 8
>16 STORED
<16 set valid_key 0 0 8
>16 STORED
<16 get ENOENT_key
>16 END
<16 connection closed.
Win.
Can someone verify that my test case is correct? If it is, I'll port
it to src/tools/regression/sockets/.
Broken or not, I don't think the behavior suggested in src/netinet/
tcp_usrreq.c:1498 is correct. A The test above write(2)'s and
read(2)'s some data successfully before testing the validity of the
half-close(2)'ed state, so I'm not worried about whether or not a SYN
has been received. I haven't looked too closely yet, but it looks
like tcp_close() needs to call to sbflush() if SO_LINGER is set (in
fact, it doesn't look like there's any special handling of linger in
tcp_subr.c, so I'm not sure if I'm glancing in the right area).
/*
* User issued close, and wish to trail through shutdown states:
* if never received SYN, just forget it. If got a SYN from peer,
* but haven't sent FIN, then go to FIN_WAIT_1 state to send peer a
FIN.
* If already got a FIN from peer, then almost done; go to LAST_ACK
* state. In all other cases, have already sent FIN to peer (e.g.
* after PRU_SHUTDOWN), and just have to play tedious game waiting
* for peer to send FIN or not respond to keep-alives, etc.
* We can let the user exit from the close as soon as the FIN is acked.
*/
Thoughts/guidance? -sc
--
Sean Chittenden
sean at chittenden.org
http://sean.chittenden.org/
More information about the freebsd-net
mailing list