aix: ignore ENOTCONN error on shutdown in uv__drain #268
Conversation
shutdown on sockets whose peer is closed causes ENOTCONN error in AIX. As we are shutting down the write part of the socket anyways, this error is of no consequence. Don't propagate, ignore instead.
|
I believe Gireesh meat to say in the last paragraph that it lets the simple/test-child-process-execsync.js pass. |
|
Yes! added a wrong test case name by mistake. Corrected now. |
|
I'm willing to believe that Linux is a bit more forgiving than AIX but what I don't understand is why the test isn't failing on BSD-like platforms. From your description of the issue it sounds like AIX has the same shutdown semantics as the BSDs and OS X: shutdown() on one half of a socket pair when the other end is closed fails with ENOTCONN. I don't want to be difficult but I have a feeling the real issue may be something else. Maybe a timing issue in the test? |
|
@bnoordhuis, thanks for your comments. In OS X as well I am able to observe the Here are the version details: Darwin Kernel Version 12.5.0 However, I see the failure pattern changes over time - probably due to the load pattern |
|
@gireeshpunathil Do you also see it with io.js? The libuv that ships with node.js v0.12 is pretty old by now and I wouldn't rule out that v0.12 itself has a couple of bugs that have been fixed in io.js. |
|
@bnoordhuis , I ran the test on the latest io.js (v1.5.2) on OS X, and found that the issue is present there too. |
|
@gireeshpunathil Thanks, I've been able to confirm the issue. I'll take a look. |
|
Continues in nodejs/node#1214. I don't think handling it in libuv is the right approach because it can mask genuine errors but it's regrettable that platform differences leak through to libuv users. I'm going to tentatively close this. |
|
|
On AIX, OS X and the BSDs, calling shutdown() on one end of a pipe when the other end has closed the connection fails with ENOTCONN. The sequential/test-child-process-execsync test failed sporadically because of a race between the parent and the child where one closed its end of the pipe before the other got around to calling shutdown() on its end of the pipe. Libuv is not the right place to handle that because it can't tell if the ENOTCONN error is genuine but io.js can. Refs: libuv/libuv#268 PR-URL: nodejs#1214 Reviewed-By: Bert Belder <[email protected]>
This is a backport of ea37ac0 Original commit message: On AIX, OS X and the BSDs, calling shutdown() on one end of a pipe when the other end has closed the connection fails with ENOTCONN. The sequential/test-child-process-execsync test failed sporadically because of a race between the parent and the child where one closed its end of the pipe before the other got around to calling shutdown() on its end of the pipe. Libuv is not the right place to handle that because it can't tell if the ENOTCONN error is genuine but io.js can. Refs: libuv/libuv#268 PR-URL: iojs#1214 Reviewed-By: Bert Belder <[email protected]> Fixes: #9444. Reviewed-By: Julien Gilli <[email protected]> PR-URL: #14480
4 participants
On AIX we were seing failures in the Node test:
simple/test-child-process-execsync.js
where the error was ENOTCONN. We identified that the
ENOTCONN a result of the call to uv_spawn in libuv.
When uv_spawn is called the child is fork/exec'd and the
parent waits for the child to terminate. Once the child
terminates it gathers the child's execution result. As
part of this process shutdown is called on the streams
which connect it to the child. At the time the parent
calls shutdown() the child will have already closed
its side of the streams. The call to shutdown is in
uv__drain() within src/unix/stream.c
Based on c test cases we ran, when run on AIX calling
shutdown on a stream for which the other side has been
closed results in the error ENOTCONN being returned.
When running the same test case on linux no error
is returned.
It appears that in uv_drain within src/unix/stream.c
the intention is to ensure we can no longer write to
the stream. If the other side is already closed this
will be the case so we believe the best course of
action is to simply ignore the ENOTCONN error in
uv__drain. And hence this pull request.
This change does not introduce any new failures when
running the tests on AIX. It also does not fix any
existing failures, but it does allow the node test
simple/test-child-process-execsync.js to pass.