[Pgpool-general] The Unplugged wire

[Tatsuo Ishii]

Daniel,

Thanks for the report!

> I spoke in another message about this problem, yet, I debugged deeper and I
> have more specific information, that, maybe, can be usefull.
> (The thread I spoke something about was:
> http://lists.pgfoundry.org/pipermail/pgpool-general/2010-February/002565.html
> )
> 
> I am working with two VB Virtual machines with CentOS 5 (i386). Running
> PostgreSQL 8.3.9 and pgpool 2.3.2.1.
> 
> The test was simple. While I was inserting values every second, I unplugged
> one of the nodes.
> health check is every second and it's timeout is 2 seconds.
> 
> In that moment all inserts stops, and pgpool waits.
> The point where it stops is:
> 
> [...]
> [pid 29444] 10:47:55.537470 rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
> [pid 29444] 10:47:55.537591 socket(PF_INET, SOCK_STREAM, IPPROTO_IP) = 9
> [pid 29444] 10:47:55.537726 setsockopt(9, SOL_TCP, TCP_NODELAY, [1], 4) = 0
> [pid 29444] 10:47:55.537886 connect(9, {sa_family=AF_INET,
> sin_port=htons(5432), sin_addr=inet_addr("192.168.1.10")}, 16) = ?
> ERESTARTSYS (To be restarted)
> [pid 29444] 10:47:56.529113 --- SIGALRM (Alarm clock) @ 0 (0) ---
> [pid 29444] 10:47:56.529235 rt_sigprocmask(SIG_SETMASK, ~[ILL TRAP ABRT BUS
> FPE SEGV CONT SYS RTMIN RT_1], NULL, 8) = 0
> [pid 29444] 10:47:56.529428 rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
> [pid 29444] 10:47:56.529602 sigreturn() = ? (mask now [])
> [pid 29444] 10:47:56.529894 connect(9, {sa_family=AF_INET,
> sin_port=htons(5432), sin_addr=inet_addr("192.168.1.10")}, 16 <unfinished
> ...>
> 
> 
> First it does a connect() wich receives de SIGALARM, and continues. But then
> it does another connect(), and this time it does not receive any SIGALARM,
> so, it waits (I think) till the system closes the connection.
> 
> After waiting (too long) it starts working again (now with the node down):
> 
> [...]
> [pid 29445] 10:49:30.273727 <... connect resumed> ) = -1 EHOSTUNREACH (No
> route to host)
> [pid 29444] 10:49:30.274739 <... connect resumed> ) = -1 EHOSTUNREACH (No
> route to host)
> [pid 29445] 10:49:30.274809 rt_sigprocmask(SIG_SETMASK, ~[ILL TRAP ABRT BUS
> FPE SEGV CONT SYS RTMIN RT_1], [], 8) = 0
> [pid 29445] 10:49:30.275057 time(NULL)  = 1267436970
> [pid 29445] 10:49:30.275202 stat64("/etc/localtime", {st_mode=S_IFREG|0644,
> st_size=2593, ...}) = 0
> [pid 29445] 10:49:30.275485 write(2, "2010-03-01 10:49:30 ERROR: pid 2"...,
> 1012010-03-01 10:49:30 ERROR: pid 29445: connect_inet_domain_socket:
> connect() failed: No route to host
> ) = 101
> [pid 29445] 10:49:30.275911 rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
> [pid 29445] 10:49:30.276062 close(7)    = 0
> [pid 29445] 10:49:30.276221 rt_sigprocmask(SIG_SETMASK, ~[ILL TRAP ABRT BUS
> FPE SEGV CONT SYS RTMIN RT_1], [], 8) = 0
> [pid 29445] 10:49:30.276389 time(NULL)  = 1267436970
> [pid 29445] 10:49:30.276715 stat64("/etc/localtime", {st_mode=S_IFREG|0644,
> st_size=2593, ...}) = 0
> [pid 29445] 10:49:30.276895 write(2, "2010-03-01 10:49:30 ERROR: pid 2"...,
> 782010-03-01 10:49:30 ERROR: pid 29445: connection to 192.168.1.10(5432)
> failed
> ) = 78
> [...]
> 
> As you can see it restarts after 1 min and a half (wich is too much). It is
> always the same (without changeing any system values)
> 
> If it is necessary I can show more debug lines.
> 
> Looking trough the source, we think, maybe it could be a problem with the
> connection being blocked. Maybe, it would be possible not to block it
> (speaking about the socket).
> We suppose something is happening in pool_connection_pool.c arround line 473
> ("connect_inet_domain_socket_by_port").
> 
> Or maybe I am doing something wrong,... does anybody else tested the
> "unpluged wire" ? Is it working?

What health_check() does here is:

start alarm (done by caller of health_check)
connect()
write()
read()
:
:

If the wire is unplugged, one of system calls will be blocked and
eventually alarm interrupt any of connect/write/read and health_check
returns with error code. Write() and read() are fine. Problem is,
connect is done by connect_inet_domain_socket_by_port, which does
retry if connect() is interrupted by a system call.

I belive what you saw was that.
> First it does a connect() wich receives de SIGALARM, and continues. But then
> it does another connect(), and this time it does not receive any SIGALARM,
> so, it waits (I think) till the system closes the connection.

The retry should be turn off if it's called from health_check(). Will
fix.

Thanks again for good testing and analysis.
--
Tatsuo Ishii
SRA OSS, Inc. Japan
English: http://www.sraoss.co.jp/index_en.php
Japanese: http://www.sraoss.co.jp