DNS设置引起的登录延迟

<p><strong>11g Network Layer Does Not Use /etc/hosts on UNIX</strong></p> <h2>Applies to: </h2> Oracle Net Services - Version: 11.1.0.6 to 11.1.0.7<br> Generic UNIX<br> <h2><a name="SYMPTOM" rel="nofollow"></a>Symptoms</h2> <p>After upgrading to 11g Oracle functionality bypasses the /etc/hosts file when resolving hostnames to IPs and instead query the DNS server. This can introduce some delay in establishing a connection to a remote host compared with 10g.</p> <p>SQL*Plus and tnsping show this behavior but this delay can appear also when opening a dblink or anytime tcpip connections are established by the Oracle Network layer.</p> <p>This happens even if  Name Server Switch configuration (nsswitch) specifies the hosts file prior to dns lookups:</p> <div>/etc/nsswitch.conf <br> hosts: files [NOTFOUND=continue] dns <br>     or only: <br> hosts: files </div> <p><br> <br> If we obtain truss/tusc trace for a SQL*Plus connection we find the following sequence of OS system calls:</p> <p>On 10g after reading the nsswitch.conf file, library "<strong>libnss_files.so</strong>" is loaded then /etc/hosts is read and the socket is opened:</p> <div> <p><strong><em>open("/etc/nsswitch.conf", O_RDONLY|0x800, 0666) = 5</em></strong> <br> ioctl(5, TCGETA, 0x9fffffffffffaca0) ERR#25 ENOTTY <br> read(5, "# \n# / e t c / n s s w i t c ".., 8192) = 92 <br> read(5, 0x60000000001e6078, 8192) = 0 <br> close(5) = 0 <br> <strong><em>open("/usr/lib/hpux64/libnss_files.so.1", O_RDONLY|0x800, 0) = 5</em></strong> <br> fstat(5, 0x9fffffffffffa720) = 0 <br> pread(5, "7fE L F 0202010101\0\0\0\0\0\0\0".., 1024, 0) = 1024 <br> stat("/usr/lib/hpux64/dpd", 0x9fffffffffff9cd0) = 0 <br> open("/usr/lib/hpux64/dpd/libnss_files.so.1.bpd", O_RDONLY|0x800, 0) ERR#2 ENOENT <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> mmap(NULL, 85872, PROT_READ|PROT_EXEC, MAP_SHARED|MAP_SHLIB, 5, 0) = 0xc0000000008d8000 <br> mmap(NULL, 3159, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_SHLIB, 5, 131072) = 0x9fffffffbf63d000 <br> close(5) = 0 <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> <strong><em>open("/etc/hosts", O_RDONLY|0x800, 0666) = 5</em></strong> <br> ioctl(5, TCGETA, 0x9fffffffffffac30) ERR#25 ENOTTY <br> read(5, "# / e t c / h o s t s \n# \n# ".., 8192) = 8192 <br> lseek(5, 18446744073709549410, SEEK_CUR) = 5986 <br> close(5) = 0 <br> <strong><em>socket(AF_INET, SOCK_STREAM, 0) = 5 <br> </em></strong>connect(5, 0x60000000001f0680, 16) = 0 <br> getsockname(5, 0x9fffffffffffb750, 0x9fffffffffffb740) = 0 <br> getsockopt(5, SOL_SOCKET, SO_SNDBUF, 0x9fffffffffffb890, 0x9fffffffffffb894) = 0 <br> getsockopt(5, SOL_SOCKET, SO_RCVBUF, 0x9fffffffffffb890, 0x9fffffffffffb894) = 0 <br> setsockopt(5, 0x6, TCP_NODELAY, 0x9fffffffffffb89c, 4) = 0</p> </div> <p> </p> <p>But on 11g after reading the nsswitch.conf  library "<strong>libnss_dns.so</strong>" is loaded then /etc/resolv.conf (which specifies available domain name servers) is read and much later a socket for TCP/IP (SOCK_STREAM) is opened. :</p> <div> <p><strong><em>open("/etc/nsswitch.conf", O_RDONLY|0x800, 0666) = 5</em> <br> </strong>ioctl(5, TCGETA, 0x9fffffffffffa360) ERR#25 ENOTTY <br> read(5, "# \n# / e t c / n s s w i t c ".., 8192) = 92 <br> read(5, 0x60000000001c9058, 8192) = 0 <br> close(5) = 0 <br> <strong><em>open("/usr/lib/hpux64/libnss_dns.so.1", O_RDONLY|0x800, 0) = 5</em></strong> <br> fstat(5, 0x9fffffffffff9de0) = 0 <br> pread(5, "7fE L F 0202010101\0\0\0\0\0\0\0".., 1024, 0) = 1024 <br> stat("/usr/lib/hpux64/dpd", 0x9fffffffffff9390) = 0 <br> open("/usr/lib/hpux64/dpd/libnss_dns.so.1.bpd", O_RDONLY|0x800, 0) ERR#2 ENOENT <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> mmap(NULL, 49440, PROT_READ|PROT_EXEC, MAP_SHARED|MAP_SHLIB, 5, 0) = 0xc00000000b054000 <br> mmap(NULL, 800, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_SHLIB, 5, 65536) = 0x9fffffffbf68d000 <br> close(5) = 0 <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> open("/test/local/oracle/11.1.0.7/lib/libdl.so.1", O_RDONLY|0x800, 0) ERR#2 ENOENT <br> open("/test/local/oracle/11.1.0.7/lib32/libdl.so.1", O_RDONLY|0x800, 0) ERR#2 ENOENT <br> getuid() = 305 (305) <br> getgid() = 303 (303) <br> open("/usr/lib/hpux64/libdl.so.1", O_RDONLY|0x800, 0) = 5 <br> fstat(5, 0x9fffffffffff9cf0) = 0 <br> read(5, "7fE L F 0202010101\0\0\0\0\0\0\0".., 64) = 64 <br> close(5) = 0 <br> socket(AF_INET, SOCK_DGRAM, 0) = 5 <br> ioctl(5, SIOCGIFNUM, 0x9fffffffffff9680) = 0 <br> ioctl(5, SIOCGIFCONF, 0x9fffffffffff9690) = 0 <br> socket(AF_INET6, SOCK_DGRAM, 0) = 6 <br> ioctl(6, SIOCGLIFNUM, 0x9fffffffffff9684) = 0 <br> ioctl(6, SIOCGLIFCONF, 0x9fffffffffff96a0) = 0 <br> ioctl(5, SIOCGIFFLAGS, 0x9fffffffffff96b0) = 0 <br> close(5) = 0 <br> close(6) = 0 <br> gettimeofday(0x9fffffffffff7dd0, NULL) = 0 <br> getpid() = 22968 (22967) <br> <strong><em>open("/etc/resolv.conf", O_RDONLY|0x800, 0666) = 5</em></strong> <br> ioctl(5, TCGETA, 0x9fffffffffff7da0) ERR#25 ENOTTY <br> read(5, "d o m a i n t e s t . c o m \n".., 8192) = 453 <br> read(5, 0x60000000001dddf8, 8192) = 0 <br> close(5) = 0 <br> ................</p> <p>.............</p> <p><strong><em>socket(AF_INET, SOCK_STREAM, 0) = 5</em></strong> <br> connect(5, 0x60000000001eba50, 16) = 0 <br> getsockname(5, 0x9fffffffffff9da0, 0x9fffffffffff94c0) = 0 <br> getsockopt(5, SOL_SOCKET, SO_SNDBUF, 0x9fffffffffffa000, 0x9fffffffffffa004) = 0 <br> getsockopt(5, SOL_SOCKET, SO_RCVBUF, 0x9fffffffffffa000, 0x9fffffffffffa004) = 0 <br> setsockopt(5, 0x6, TCP_NODELAY, 0x9fffffffffffa00c, 4) = 0 <br> </p> </div> <p><br>   </p> <h2><a name="CHANGE" rel="nofollow"></a>Changes</h2> <p>Nothing was changed in the configuration of the OS, only the upgrade from Oracle 10g to 11g was done. </p> <h2><a name="CAUSE" rel="nofollow"></a>Cause</h2> <p>What was changed between the two is the way Oracle resolves hostnames to IPs, more specifically the system call used to do that.</p> <p>Oracle 11g use now <strong><em>getaddrinfo()</em></strong> while 10g used <strong><em>gethostbyname().</em></strong></p> <p>These system functions requires different configuration in /etc/nsswitch.conf.</p> <p><strong><em>gethostbyname()</em></strong> require the use of keyword "hosts" while <strong><em>getaddrinfo()</em></strong> the use the keyword "ipnodes"</p> <p> </p> <p> </p> <p>Notes:<br> Even though this has only been observed on HP-UX and Solaris, this may be UNIX generic.<br> With Solaris, ipnodes has a different meaning (specify a file for IPV6 addresses resolution, gethostbyname and getaddrinfo both use hosts or ipnodes file). <br> Linux on the other hand does not use ipnodes in nsswitch.conf</p> <h2><a name="FIX" rel="nofollow"></a>Solution</h2> <p>Add a line in the /etc/nsswitch.conf file similar to the following:<br>     ipnodes: files [NOTFOUND=continue] dns</p> <p>The line starting with keyword "hosts" must not be deleted.</p> <p>This way calls made by getaddrinfo() will search first in /etc/hosts then, if the name is not found, will contact the dns server. <br> Thus there will be no connection delay for any lookup of host names existing in the local /etc/hosts file.</p> <p> </p>

<p><strong>Sqlnet connection via TCP Hangs for about 10seconds Before Connection is Established</strong></p> <h2>Applies to: </h2> Oracle Server - Enterprise Edition - Version: 10.1.0.2 to 11.2.0.1.0 - Release: 10.1 to 11.2<br> Information in this document applies to any platform.<br> <h2><a name="SYMPTOM" rel="nofollow"></a>Symptoms</h2> 10 nodes RAC cluster, sqlnet connection via TCP consistently takes 10 seconds before the connection with the server is finally established.<br> <br> <div> $date;sqlplus apps/xxx@dwnodea @time.sql<br> Tue Feb  2 17:00:39 CST 2010<br> <br> SQL*Plus: Release 11.1.0.7.0 - Production on Tue Feb  2 17:00:39 2010<br> <br> Copyright (c) 1982, 2008, Oracle. All rights reserved.<br> <br> Connected to:<br> Oracle Database 11g Enterprise Edition Release 11.1.0.7.0 - 64bit Production<br> With the Partitioning, Real Application Clusters, OLAP and Real Application Testing options<br> <br> TO_CHAR(SYSDATE,'YYYY-MM-DDHH24:MI:SS')<br> ------------------------------------------------------------------------------------------------------------------------<br> 2010-02-02 17:00:49</div> <br> Sqlnet tracing (both server and client) against sqlplus connection with level 16 shows:<br> <br> <div> Server side trace:<br> 2010-02-02 17:00:39.432277 : nttcnp:exit<br> <strong>2010-02-02 17:00:39.432308 : snlinGetAddrInfo:entry</strong><br> <br> *** 2010-02-02 17:00:49.440<br> <strong>2010-02-02 17:00:49.440655 : snlinGetAddrInfo:getaddrinfo() failed with error -2</strong><br> 2010-02-02 17:00:49.440736 : snlinGetAddrInfo:exit<br> 2010-02-02 17:00:49.440751 : nttcon:exit<br> 2010-02-02 17:00:49.440767 : nsopen:transport is open<br> 2010-02-02 17:00:49.440785 : nsoptions:entry<br> 2010-02-02 17:00:49.440802 : nsoptions:lcl[0]=0x0, lcl[1]=0x0, gbl[0]=0x0, gbl[1]=0x0, cha=0x0</div> <br> <br> <div> Client side trace:<br> 2010-02-02 17:00:39.418598 : nscon:recving a packet<br> 2010-02-02 17:00:39.418714 : nsprecv:entry<br> 2010-02-02 17:00:39.418833 : nsprecv:reading from transport...<br> <strong>2010-02-02 17:00:39.418949 : nttrd:entry<br> 2010-02-02 17:00:49.444643 : nttrd:socket 12 had bytes read=8</strong><br> 2010-02-02 17:00:49.444832 : nttrd:exit</div> <br> <br> <h2><a name="CHANGE" rel="nofollow"></a>Changes</h2> This is a new installation. <h2><a name="CAUSE" rel="nofollow"></a>Cause</h2> This issue is caused by invalid nameserver entry in /etc/resolv.conf.<br> <br> Examine system configuration files show:<br> <br> <div>resolv.conf <br> --------<br> nameserver 192.168.1.20<br> nameserver 192.168.4.12<br> <br> nsswitch.conf<br> ----------<br> #hosts: db files ldap nis dns<br> hosts: files dns</div> <br> These are incorrect nameserver per customer configuration, they do not exist. Once they are removed, there is no longer any delay in sqlnet connection.<br> <h2><a name="FIX" rel="nofollow"></a>Solution</h2> 1. Remove incorrect setting for nameserver in /etc/resolv.conf if DNS is not in use.<br> or<br> 2. Consult with the network administrator to provide correct nameserver IP address if DNS is in use.

========================================================================= Interim Patch for Base Bug: 9593134 ========================================================================= Date: Tue Aug 10 18:18:48 2010 ------------------------------------------------------------------------- Platform Patch for : Linux-x86 Product Patched : SQLNET Product Version # : 11.1.0.7.0 RAC Rolling Installable : YES Bugs Fixed by this patch: ------------------------- 9593134: DNS OR NIS MIS-CONFIGURATION CAN CAUSE SLOW DATABASE CONNECTS Hdr: 8307164 11.1.0.7 NET 11.1.0.7 PRODID-115 PORTID-197 Abstract: TNSPING 11G USING DNS AND NOT HOSTS FILE *** 03/04/09 05:24 am *** TAR: ---- Problem: -------- Customer has 2 oracle homes on the same HP-UX server, one 10g and one 11.1.0.7. He execute tnsping <alias> in both versions, and truss them truss -o out tnsping <alias> In 10g truss shows that the hostname from the connect descriptor is resolved using /etc/hosts file but in 11g the same hostname is resolved using dns. The same behavior is observed for sqlplus. Because of this connections to the database ar slow because of the slow hostname resolution using the dns. Versions: --------- Diagnostic Analysis: -------------------- nsswitch.conf ============== hosts: files [NOTFOUND=continue] dns Reproducibility: ---------------- on customer site, also I observered same behavior on internal HP-UX server with 11g database Test Case: ---------- Workaround: ----------- none

Hdr: 5047830 9.2.0.6 NET 9.2.0.6 PRODID-115 PORTID-59 2728394 Abstract: TNSPING HANGS AND CONSUMES MEMORY UUNTIL THE SERVER LOCKS UP Problem: -------- Problem Description ~~~~~~~~~~~~~~~~~~~~ Firstly, the problem with tnsping hanging when locating a bad tnsnames alias was reported in Bug2728394 "TNSPING HANGS AND DOES NOT RETURN ERROR WHEN SERVICE NAME IS WRONG" This was shown for all 9.2.0.x versions. HOWEVER, the customer has responded saying that the "workaround" of making sure the tnsnames.ora aliases are not corrupt is unacceptable for the following severe reason: If a tnsping is inadvertantly issued on a production server against a "bad" tnsnames.ora alias, the tnsping command hangs, BUT memory resources leak and eventually, after a few minutes (depending on the resources available) the OS runs out of memory and causes a production outage. The solution is to kill any hanging tnsping attempt but this could happen without the DBA's knowledge until the situation is severe. Versions: --------- SQL*Net / Net Services 9.2.0.6 and 9.2.0.7 Tested against HP-UX 11 (customer's platform) = 9.2.0.6 Tested against Solaris 2.9 = 9.2.0.7. Same results each time. Diagnostic Analysis: -------------------- 1. Manually create a "bad" tnsnames.ora alias. 2. Run tnsping against the alias. 3. Run 'top' or some other tool to check memory leak. 4. Kill the hung process Reproducibility: ---------------- Everytime in-house and customer's site. Test Case: ---------- Testcase uploaded as testcase.txt Workaround: ----------- Kill the hanging tnsping process. However, not easy if there is no knowledge of the tnsping that was run.

感谢ML，学习了，真长知识