What does getaddrinfo do?

The following C program calls getaddrinfo("google.com", ...), a function from sys/socket.h. On the face of it, getaddrinfo is used to do DNS lookups.

#include <stdio.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <netdb.h>
#include <arpa/inet.h>
int main(void) {
  struct addrinfo* addr;
  int result = getaddrinfo("google.com", NULL, NULL, &addr);
  if (result != 0) {
    printf("Error from getaddrinfo: %s\n", gai_strerror(result));
    return 1;
  }
  struct sockaddr_in* internet_addr = (struct sockaddr_in*) addr->ai_addr;
  printf("google.com is at: %s\n", inet_ntoa(internet_addr->sin_addr));
  return 0;
}

For me, this program prints:

$ cc main.c
$ ./a.out
google.com is at: 216.58.208.142

Great! But what is this program doing to determine that google.com can be found at 216.58.208.142? You might imagine that getaddrinfo is a wrapper around a DNS request, and that if we snooped on the traffic, we’d see two UDP packets, one for request and one for response:

$ sudo tcpdump -v 'host 10.0.2.3'
17:37:39.915177 IP (tos 0x0, ttl 64, id 21777, offset 0, flags [DF], proto UDP (17), length 56)
    vagrant-ubuntu-trusty-64.55983 > 10.0.2.3.domain: 47922+ A? google.com. (28)
17:37:39.942126 IP (tos 0x0, ttl 64, id 2975, offset 0, flags [none], proto UDP (17), length 72)
    10.0.2.3.domain > vagrant-ubuntu-trusty-64.55983: 47922 1/0/0 google.com. A 216.58.208.142 (44)

We do see those packets: a request for an A record for google.com, and its response. But that’s not all we see. The following is the real output, showing an additional request/response for an AAAA record. This asks for an IPv6 address, and the server replies with 2a00:1450:4009:803::200e.

$ sudo tcpdump -v 'host 10.0.2.3'
17:37:39.915177 IP (tos 0x0, ttl 64, id 21777, offset 0, flags [DF], proto UDP (17), length 56)
    vagrant-ubuntu-trusty-64.55983 > 10.0.2.3.domain: 47922+ A? google.com. (28)
17:37:39.915667 IP (tos 0x0, ttl 64, id 21778, offset 0, flags [DF], proto UDP (17), length 56)
    vagrant-ubuntu-trusty-64.55983 > 10.0.2.3.domain: 61082+ AAAA? google.com. (28)
17:37:39.942126 IP (tos 0x0, ttl 64, id 2975, offset 0, flags [none], proto UDP (17), length 72)
    10.0.2.3.domain > vagrant-ubuntu-trusty-64.55983: 47922 1/0/0 google.com. A 216.58.208.142 (44)
17:37:39.945624 IP (tos 0x0, ttl 64, id 2977, offset 0, flags [none], proto UDP (17), length 84)
    10.0.2.3.domain > vagrant-ubuntu-trusty-64.55983: 61082 1/0/0 google.com. AAAA 2a00:1450:4009:803::200e (56)

This isn’t so bad. But oh boy, is getaddrinfo more complex than this! We’ve seen what getaddrinfo does using tcpdump; now let’s try another angle: strace. This shows all system calls that our program makes:

$ strace ./a.out
execve("./a.out", ["./a.out"], [/* 20 vars */]) = 0
brk(0)                                  = 0x1fbf000
...
...

I initially dumped the entire strace output into this blog post, but then decided it was too large and you’d get put off. The getaddrinfo function call alone causes over 100 system calls! Files are read, dynamic libraries are loaded, obscure sockets are opened to the kernel, obscure sockets are opened to non-existent daemons, multiple times ... So let’s break it down, and I’ll try to reduce it to the most relevant system calls. Stay with me as I attempt to trace the crazy path of address lookup on Linux.

But first, let’s see the money shot, the system calls which make a DNS request to the DNS server and read the response:

socket(PF_INET, SOCK_DGRAM|SOCK_NONBLOCK, IPPROTO_IP) = 3
connect(3, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("10.0.2.3")}, 16) = 0
poll([{fd=3, events=POLLOUT}], 1, 0)    = 1 ([{fd=3, revents=POLLOUT}])
sendmmsg(3, {{{msg_name(0)=NULL, msg_iov(1)=[{"\244>\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\1\0\1", 28}], msg_controllen=0, msg_flags=0}, 28}, {{msg_name(0)=NULL, msg_iov(1)=[{"\332\263\1\0\0\1\0\0\0\0\0\0\6google\3com\0\0\34\0\1", 28}], msg_controllen=0, msg_flags=0}, 28}}, 2, MSG_NOSIGNAL) = 2
poll([{fd=3, events=POLLIN}], 1, 5000)  = 1 ([{fd=3, revents=POLLIN}])
ioctl(3, FIONREAD, [44])                = 0
recvfrom(3, "\244>\201\200\0\1\0\1\0\0\0\0\6google\3com\0\0\1\0\1\300\f\0\1"..., 2048, 0, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("10.0.2.3")}, [16]) = 44
poll([{fd=3, events=POLLIN}], 1, 4981)  = 1 ([{fd=3, revents=POLLIN}])
ioctl(3, FIONREAD, [56])                = 0
recvfrom(3, "\332\263\201\200\0\1\0\1\0\0\0\0\6google\3com\0\0\34\0\1\300\f\0\34"..., 65536, 0, {sa_family=AF_INET, sin_port=htons(53), sin_addr=inet_addr("10.0.2.3")}, [16]) = 56
close(3)                                = 0

The above system calls correlate with the UDP packet dump. Two packets are sent with one sendmmsg, then the two responses are read with recvfrom.

(Notice in passing the calls to poll, blocking the entire process for up to 5 seconds waiting for the DNS response. In a program with an event loop or “green threads”, this can be disastrous. getaddrinfo does not play nicely with your event loop!)

But getaddrinfo does a lot before these system calls, and it does quite a bit after them, too. You see the UDP “connection” is to a DNS server at 10.0.2.3. But how did getaddrinfo know about 10.0.2.3? It’s not a function argument. Just before the DNS request, the process makes these system calls:

open("/etc/resolv.conf", O_RDONLY|O_CLOEXEC) = 3
read(3, "# Dynamic resolv.conf(5) file fo"..., 4096) = 182

getaddrinfo gets the DNS server’s IP address from a file at /etc/resolv.conf! Here’s my /etc/resolv.conf:

$ cat /etc/resolv.conf
# Dynamic resolv.conf(5) file for glibc resolver(3) generated by resolvconf(8)
#     DO NOT EDIT THIS FILE BY HAND -- YOUR CHANGES WILL BE OVERWRITTEN
nameserver 10.0.2.3
search lan

As the comments in the file imply, there’s a complex process which generates the resolv.conf. Ultimately, the IP address comes from DHCP, but this is another story.

But host addresses information doesn’t only come from DNS! There’s another source of information on UNIX systems: the file /etc/hosts. Here’s mine:

$ cat /etc/hosts
127.0.0.1 localhost

# The following lines are desirable for IPv6 capable hosts
::1 ip6-localhost ip6-loopback
fe00::0 ip6-localnet
ff00::0 ip6-mcastprefix
ff02::1 ip6-allnodes
ff02::2 ip6-allrouters
ff02::3 ip6-allhosts

If I add the line 1.2.3.4 google.com, the output of my program changes:

$ ./a.out
google.com is at: 1.2.3.4

getaddrinfo gets info from /etc/hosts in the obvious way, by reading that file in full every time you call it:

open("/etc/hosts", O_RDONLY|O_CLOEXEC)  = 3
read(3, "127.0.0.1 localhost\n\n# The follo"..., 4096) = 221

For some time I thought this was the whole story, and that the C standard library has hardcoded knowledge of /etc/hosts and /etc/resolv.conf. But in fact, the C standard library doesn’t have knowledge of either of these files! This knowledge is actually coded in some shared dynamic libraries. The file /etc/hosts is known by /lib/x86_64-linux-gnu/libnss_files.so.2, and /etc/resolv.conf is known by /lib/x86_64-linux-gnu/libnss_dns.so.2. Those libraries are loaded at runtime by our process, like this:

open("/lib/x86_64-linux-gnu/libnss_dns.so.2", O_RDONLY|O_CLOEXEC) = 3
read(3, "\177ELF\2\1\1\0\0\0\0\0\0\0\0\0\3\0>\0\1\0\0\0\0\21\0\0\0\0\0\0"..., 832) = 832
fstat(3, {st_mode=S_IFREG|0644, st_size=22952, ...}) = 0
mmap(NULL, 2117896, PROT_READ|PROT_EXEC, MAP_PRIVATE|MAP_DENYWRITE, 3, 0) = 0x7f5216c88000
mprotect(0x7f5216c8d000, 2093056, PROT_NONE) = 0
mmap(0x7f5216e8c000, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_FIXED|MAP_DENYWRITE, 3, 0x4000) = 0x7f5216e8c000

You might think, “Jim, you’re being picky. This is still the C standard library; it just lazily loads its files and dns libraries.” But this is not how it works! Although we think of getaddrinfo as “do DNS lookup”, its man page never even mentions DNS!:

Given node and service, which identify an Internet host and a service, getaddrinfo() returns one or more addrinfo structures, each of which contains an Internet address that can be specified in a call to bind(2) or connect(2).

getaddrinfo doesn’t know anything about files, DNS, or any other way to find the address for a host. Instead, getaddrinfo gets a list of these “sources” at runtime from another file, /etc/nsswitch.conf, the “Name Service Switch”. Here’s some of mine:

$ cat /etc/nsswitch.conf
passwd:         compat
hosts:          files myhostname dns
networks:       files

Notice the line hosts: files myhostname dns. This says, “to find a host, first ask the library libnss_files.so. If that fails, ask the library libnss_myhostname.so. Finally, ask the library libnss_dns.so.” The C standard library interpolates files, dns and so on into the pattern libnss_%s.so to find the libraries. As such, you could write a new library libnss_imfeelinglucky, and change your nsswitch.conf to hosts: imfeelinglucky. Enjoy the chaos.

You might think we’re done. Not yet! Before getaddrinfo does any of this, we have these system calls:

socket(PF_LOCAL, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0) = 3
connect(3, {sa_family=AF_LOCAL, sun_path="/var/run/nscd/socket"}, 110) = -1 ENOENT (No such file or directory)
close(3)                                = 0

What is /var/run/nscd/socket ..? Linux tells us, with ENOENT, that I don’t have that file! What is this supposed to be? As Google will tell you, this is a socket to talk to the Northern School of Contemporary Dance. But before your process can go to class, you have to install the daemon:

$ sudo apt-get install nscd
...
Setting up nscd (2.19-0ubuntu6.14) ...
 * Starting Name Service Cache Daemon

Sorry, nscd is actually the “name service cache daemon”, “a daemon that provides a cache for the most common name service requests”. After installing it, the daemon starts, and your process can dance:

socket(PF_LOCAL, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0) = 4
connect(4, {sa_family=AF_LOCAL, sun_path="/var/run/nscd/socket"}, 110) = 0
sendto(4, "\2\0\0\0\r\0\0\0\6\0\0\0hosts\0", 18, MSG_NOSIGNAL, NULL, 0) = 18
poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN|POLLERR|POLLHUP}])
recvmsg(4, 0x7fff7a50f5f0, MSG_CMSG_CLOEXEC) = -1 ECONNRESET (Connection reset by peer)
close(-1)                               = -1 EBADF (Bad file descriptor)
close(4)                                = 0
close(3)                                = 0
socket(PF_LOCAL, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0) = 3
connect(3, {sa_family=AF_LOCAL, sun_path="/var/run/nscd/socket"}, 110) = -1 ECONNREFUSED (Connection refused)
close(3)

Oh, wait, no. The daemon reset the connection. Then a bug in glibc closes the file descriptor -1. Then glibc tries to connect to NSCD again, just to make sure. (Oh, glibc doesn’t just do that in this circumstance. glibc tries twice, every time.) So what happened?

$ sudo tail -2 /var/log/kern.log
Feb  3 19:36:17 vagrant-ubuntu-trusty-64 kernel: [11799.496494] nscd[3677]: segfault at 43c000010 ip 00007fba29180753 sp 00007fba1e4741f0 error 4 in nscd[7fba2916c000+25000]
Feb  3 19:39:46 vagrant-ubuntu-trusty-64 kernel: [12008.644917] nscd[3758]: segfault at 0 ip 00007ff37679cdfa sp 00007ff36ce901e8 error 4 in libc-2.19.so[7ff376714000+1be000]

Turns out that nscd segfaulted. As everyone on the internet will tell you, nscd is “shit”, “unstable”, and “badly designed”. But no worries: as part of the C standard library, nscd is not this program but a standard protocol, part of the standard UNIX specification, with many alternative implementations!

At least, that’s what I thought. As it turns out, the source of nscd is buried in glibc. (Despite this program not being part of libc!) The protocol for nscd is not specified, and there are virtually no alternative implementations! nscd is not part of the C standard library. If you use a different implementation like musl-libc, your programs won’t look for the name server cache daemon.

Rant over, back on track. Before glibc looks for nscd, it does some even more mysterious stuff:

socket(PF_NETLINK, SOCK_RAW, 0)         = 3
bind(3, {sa_family=AF_NETLINK, pid=0, groups=00000000}, 12) = 0
getsockname(3, {sa_family=AF_NETLINK, pid=2166, groups=00000000}, [12]) = 0
sendto(3, "\24\0\0\0\26\0\1\3\1\345uZ\0\0\0\0\0\0\0\0", 20, 0, {sa_family=AF_NETLINK, pid=0, groups=00000000}, 12) = 20
recvmsg(3, {msg_name(12)={sa_family=AF_NETLINK, pid=0, groups=00000000}, msg_iov(1)=[{"D\0\0\0\24\0\2\0\1\345uZv\10\0\0\2\10\200\376\1\0\0\0\10\0\1\0\177\0\0\1"..., 4096}], msg_controllen=0, msg_flags=0}, 0) = 148
recvmsg(3, {msg_name(12)={sa_family=AF_NETLINK, pid=0, groups=00000000}, msg_iov(1)=[{"@\0\0\0\24\0\2\0\1\345uZv\10\0\0\n\200\200\376\1\0\0\0\24\0\1\0\0\0\0\0"..., 4096}], msg_controllen=0, msg_flags=0}, 0) = 128
recvmsg(3, {msg_name(12)={sa_family=AF_NETLINK, pid=0, groups=00000000}, msg_iov(1)=[{"\24\0\0\0\3\0\2\0\1\345uZv\10\0\0\0\0\0\0", 4096}], msg_controllen=0, msg_flags=0}, 0) = 20

Our process uses a PF_NETLINK socket to talk to the kernel. But what is it saying? The glibc source says it’s trying to find “information about what interfaces are available. Also determine whether we have IPv4 or IPv6 interfaces or both.” It’s not clear why our program needs to know what interfaces are available.

BUT WAIT THERE’S STILL MORE! After our process has its DNS responses, it does more work. It starts by reading /etc/gai.conf, the “Configuration for getaddrinfo(3).” The function call has its very own configuration file! Luckily, mine is only comments.

Once getaddrinfo has the addresses, you’d think it would just return them. But first, it makes some tests on those addresses, opening sockets to them and connecting:

socket(PF_INET, SOCK_DGRAM, IPPROTO_IP) = 3
connect(3, {sa_family=AF_INET, sin_port=htons(0), sin_addr=inet_addr("216.58.208.142")}, 16) = 0
getsockname(3, {sa_family=AF_INET, sin_port=htons(59663), sin_addr=inet_addr("10.0.2.15")}, [16]) = 0
close(3)                                = 0
socket(PF_INET6, SOCK_DGRAM, IPPROTO_IP) = 3
connect(3, {sa_family=AF_INET6, sin6_port=htons(0), inet_pton(AF_INET6, "2a00:1450:4009:803::200e", &sin6_addr), sin6_flowinfo=0, sin6_scope_id=0}, 28) = -1 ENETUNREACH (Network is unreachable)
close(3)                                = 0

One last thing. You might think that getaddrinfo caches answers, so subsequent calls aren’t so expensive. It does not! It does this entire procedure every time!

Discussion on Hacker News.
Tagged #programming, #c, #posix.

Similar posts

More by Jim

Want to build a fantastic product using LLMs? I work at Granola where we're building the future IDE for knowledge work. Come and work with us! Read more or get in touch!

This page copyright James Fisher 2018. Content is not associated with my employer. Found an error? Edit this page.