[
https://issues.apache.org/jira/browse/THRIFT-5186?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17129807#comment-17129807
]
Jens Geyer commented on THRIFT-5186:
------------------------------------
Original comment from https://github.com/apache/thrift/pull/2151
My previous patch (9b9567b23378c821b460cfe54b70b9d189bf194d) has exposed
an issue shared by TServerSocket and TNonblockingServerSocket:
the results of getaddrinfo() call aren't used "in full" -- just a single
address is picked, and then bind() is retried in a loop on that single address.
This leads to poor results if we start varying the network conditions.
Like I show in the Jira issue, this is normal:
[root@04dd07b70038 /]# ping -6 localhost
ping: connect: Cannot assign requested address
-- same with `ping -6 $(hostname)` -- a hostname may resolve to IPv6 address
which we might not be able to connect to; and vice-versa, connecting to
IPv4 addresses may fail on IPv6-only systems.
The solution is to iterate over what getaddrinfo() returns while retrying
failed bind(). This is what this patch implements. Server-side analogue
of this behavior of curl:
> curl -v localhost:8001/index.do
* Trying ::1:8001...
* connect to ::1 port 8001 failed: Connection refused
* Trying 127.0.0.1:8001...
* Connected to localhost (127.0.0.1) port 8001 (#0)
> GET /index.do HTTP/1.1
[...]
To achieve that, I throw away the TGetAddrInfoWrapper, and roll a proper one.
The bind-retrying loop had to be restructured: since sa_family will vary
from retry to retry, the socket() call has to be within the loop.
All the setsockopt() business factored away into side-methods: ~hundred
repetitive #ifdef-rich lines in the middle of important loop do get in the way.
This patch has quite a lot of code movement; git config diff.colormoved zebra
is strongly advised to the reader.
Tested manually on Linux, by running CMake's `make test` in 3 Docker
environments:
1) loopback-only (--net=none)
2) classic IPv4-only bridge
3) dual IPv6-IPv4 bridge
The only failing test was 88 - PythonTestSSLSocket, which currently also
fails on master for me (due to NULL cipher being unexpectedly accepted).
> AI_ADDRCONFIG: Thrift libraries crash with localhost-only network.
> ------------------------------------------------------------------
>
> Key: THRIFT-5186
> URL: https://issues.apache.org/jira/browse/THRIFT-5186
> Project: Thrift
> Issue Type: Bug
> Components: C++ - Library, Delphi - Library, Python - Library
> Affects Versions: 0.13.0
> Environment: Red Hat Enterprise Linux 8.0
> Reporter: Max
> Assignee: Max
> Priority: Major
> Labels: getaddrinfo, localhost, sockets
> Fix For: 0.14.0
>
> Attachments:
> 0001-THRIFT-5186-Dont-pass-AI_ADDRCONFIG-to-getaddrinfo.patch
>
> Time Spent: 6h 20m
> Remaining Estimate: 0h
>
> THRIFT-2539 has been reported, and fixed — but for win32 only, for no
> apparent reason. The exact same problem reproduces on POSIX.
> Namely, when no network interfaces besides {{lo}} (the 127.0.0.1 loopback
> interface) are up, C++ and Python apps linked with Thrift-generated code,
> both clients and servers — *crash by throwing an exception*. Even when the
> intention is exactly to run them on localhost only.
> This happens because Thrift library code for TSocket, TServerSocket,
> TNonblockingServerSocket calls
> [{{getaddrinfo()}}|http://man7.org/linux/man-pages/man3/getaddrinfo.3.html]
> to resolve target hostname to connect to/listen on, into concrete IP address
> (v4 or v6, whichever the system is configured for). To that call, it *passes
> the {{AI_ADDRCONFIG}} hint* which effectively turns a localhost-only
> situation into:
> {quote}{{Could not resolve host for client socket.}}
> {quote}
> and into this (server-side):
> {code:java}
> гру 23 13:52:13 localhost.localdomain systemd[1]: db_cache.service: Main
> process exited, code=dumped, status=6/ABRT
> гру 23 13:52:13 localhost.localdomain systemd[1]: db_cache.service: Failed
> with result 'core-dump'.
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: Thrift: Mon Dec 23
> 13:52:15 2019 TSocket::open() getaddrinfo() <Host: 127.0.0.1 Port:
> 1302>Address family for hostname not supported
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: Thrift: Mon Dec 23
> 13:52:15 2019 TSocket::open() getaddrinfo() <Host: 127.0.0.1 Port:
> 8345>Address family for hostname not supported
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: Thrift: Mon Dec 23
> 13:52:15 2019 TNonblocking: using dedicated listener thread, io threads: 16
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: Thrift: Mon Dec 23
> 13:52:15 2019 getaddrinfo -9: Address family for hostname not supported
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: terminate called after
> throwing an instance of 'apache::thrift::transport::TTransportException'
> гру 23 13:52:17 localhost.localdomain db_cache[12912]: what(): Could not
> resolve host for server socket.
> {code}
> I fail to understand the original reason to pass that {{AI_ADDRCONFIG}} hint.
> It shouldn't be there as I see it.
> Further, since Thrift 0.9.2, windows builds of thrift apps don't pass that
> hint anymore (see THRIFT-2539), and it seems to be okay.
> For comprehension, I'm attaching a sample patch to remove {{AI_ADDRCONFIG}}
> from {{lib/cpp}} and {{lib/py}}. The main change will be landing via GitHub,
> per Thrift's contribution process, so please follow there too.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
