Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

net: connect after polling initialization #8426

Closed
gopherbot opened this issue Jul 26, 2014 · 23 comments

Comments

Projects
None yet
5 participants
@gopherbot
Copy link

commented Jul 26, 2014

by jason@eggnet.com:

I have checked the latest source, this bug applies as of the current master branch.

go is calling epoll_ctl ADD before connect which I believe is incorrect. If another
thread is in epoll_wait, that thread can (will?) get an event with EPOLLOUT|EPOLLHUP
flags before connect is called, which is what has been happening.

In general this appears to be a general flaw with all forms of polling for all operating
systems, but I have only performed extensive testing with linux amd64 / epoll. If I knew
of another popular program that managed the same epoll file descriptor in multiple
threads I would include information about that here, but I do not. nginx does call
epoll_ctl_add prior to connect but it is single threaded. When using multiple processes
each nginx process has one thread with its own instance of epoll. Since there is no
other thread in epoll_wait, the order of operations I suspect is irrelevant for nginx.

The call chain for linux for adding a new outbound socket to epoll is shown. The fd.dial
function is deciding to call fd.init before connecting. I'm not necessarily saying the
fix is as simple as altering the order of operations in fd.dial. If connect always
responded with EINPROGRESS it might be that simple. But apparently certain operating
systems will return from connect immediately with no error when connecting to localhost.
If that happens, then there are no pending operations on the new socket. It doesn't need
to be monitored by epoll. Therefore some care has to be taken, I think, in terms of
dealing with the pd and fd state.

d.Dial : dial.go
dialSingle : dial.go
dialTCP : tcpsock_posix.go
socket : sock_posix.go
fd.dial : sock_posix.go
fd.init : fd_unix.go
fd.pd.Init : fd_poll_runtime.go
runtime_pollOpen : znetpoll_linux_amd64.c

I wrote a quick test in C, relevant syscalls highlighted below. The test was to make
sure that epoll_wait was able to see the connect event even after a substantial delay
between connect and epoll_ctl.

1406334260.600586 socket(PF_INET, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, IPPROTO_IP) =
4 <0.000013>
1406334260.600629 setsockopt(4, SOL_SOCKET, SO_BROADCAST, [1], 4) = 0 <0.000006>
1406334260.600671 connect(4, {sa_family=AF_INET, sin_port=htons(3456),
sin_addr=inet_addr("69.28.182.189")}, 16) = -1 EINPROGRESS (Operation now in
progress) <0.000054>
1406334260.600808 rt_sigprocmask(SIG_BLOCK, [CHLD], [], 8) = 0 <0.000005>
1406334260.600859 rt_sigaction(SIGCHLD, NULL, {SIG_DFL, [], 0}, 8) = 0 <0.000005>
1406334260.600913 rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0 <0.000004>
1406334260.600946 nanosleep({5, 0}, 0x7fff83894450) = 0 <5.000141>
1406334265.601188 epoll_ctl(3, EPOLL_CTL_ADD, 4, {EPOLLIN|EPOLLOUT|EPOLLET|0x2000,
{u32=4, u64=18023070048452612}}) = 0 <0.000010>
1406334265.601237 epoll_wait(3, {{EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000, {u32=4,
u64=18023070048452612}}}, 1, 10000) = 1 <0.000005>
1406334265.601310 getsockopt(4, SOL_SOCKET, SO_ERROR, [111], [4]) = 0 <0.000006>
1406334265.601350 close(4)              = 0 <0.000018>


Here's a clearer picture of what is happening in go:

http://play.golang.org/p/wXceILeetG

This is with go's internal resolver.

Running the trace like this:

strace -ttt -T -ff -o a ./test hostname:3456

Find the write that returns ECONNREFUSED.

Determine which file descriptor it is, then determine which u64 ID is used with it in
epoll_ctl. This should show all relevant syscalls that relate to the file descriptor or
u64 ID.

The spurious epoll_wait event fires between epoll_ctl and connect.

egrep '[(, ]215($|[), ])|139969772241984' a.* | perl -pe 's/^a.(\d+):(\S+)/$2 $1/' | sort

1406320197.172769 26587 getsockopt(439, SOL_SOCKET, SO_ERROR, [111], [4]) = 0
<0.000005>
1406320197.172794 26587 epoll_ctl(4, EPOLL_CTL_DEL, 439, {0, {u32=4213032,
u64=4213032}}) = 0 <0.000006>
1406320197.172817 26587 close(439)            = 0 <0.000010>
1406320197.507617 26587 socket(PF_INET, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK,
IPPROTO_IP) = 439 <0.000010>
1406320197.507645 26587 setsockopt(439, SOL_SOCKET, SO_BROADCAST, [1], 4) = 0
<0.000006>
1406320197.507672 26587 epoll_ctl(4, EPOLL_CTL_ADD, 439,
{EPOLLIN|EPOLLOUT|EPOLLET|0x2000, {u32=2368844064, u64=139971058053408}}) = 0
<0.000006>
1406320197.507823 26588 epoll_wait(4, {{EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368724616, u64=139971057933960}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368839488, u64=139971058048832}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368774368, u64=139971057983712}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368821712, u64=139971058031056}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368839136, u64=139971058048480}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368757648, u64=139971057966992}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368839312, u64=139971058048656}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368757296, u64=139971057966640}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368757120, u64=139971057966464}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368757472, u64=139971057966816}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368756944, u64=139971057966288}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368722504, u64=139971057931848}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368722328, u64=139971057931672}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368791088, u64=139971058000432}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790912, u64=139971058000256}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368780528, u64=139971057989872}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790560, u64=139971057999904}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790384, u64=139971057999728}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790736, u64=139971058000080}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790032, u64=139971057999376}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368790208, u64=139971057999552}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368789680, u64=139971057999024}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368789328, u64=139971057998672}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368789856, u64=139971057999200}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368789152, u64=139971057998496}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368788976, u64=139971057998320}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368789504, u64=139971057998848}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368761696, u64=139971057971040}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368788800, u64=139971057998144}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368788624, u64=139971057997968}}, {EPOLLOUT|EPOLLHUP, {u32=2368844064,
u64=139971058053408}}}, 128, 0) = 31 <0.000010>
1406320197.507942 26587 connect(439, {sa_family=AF_INET, sin_port=htons(3456),
sin_addr=inet_addr("x")}, 16) = -1 EINPROGRESS (Operation now in progress)
<0.000031>
1406320197.508003 26587 getsockopt(439, SOL_SOCKET, SO_ERROR, [0], [4]) = 0
<0.000005>
1406320197.508081 26587 getsockname(439, {sa_family=AF_INET, sin_port=htons(36051),
sin_addr=inet_addr("x")}, [16]) = 0 <0.000005>
1406320197.508109 26587 getpeername(439, 0x7f4d8cdb4560, [112]) = -1 ENOTCONN (Transport
endpoint is not connected) <0.000005>
1406320197.508141 26587 setsockopt(439, SOL_TCP, TCP_NODELAY, [1], 4) = 0
<0.000006>
1406320197.508262 26587 write(439, "hi", 2)   = -1 EAGAIN (Resource
temporarily unavailable) <0.000006>
1406320197.518050 26588 epoll_wait(4, {{EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368843888, u64=139971058053232}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368843712, u64=139971058053056}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368843536, u64=139971058052880}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368844768, u64=139971058054112}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368754128, u64=139971057963472}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368806224, u64=139971058015568}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368782992, u64=139971057992336}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368761168, u64=139971057970512}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368806048, u64=139971058015392}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368833504, u64=139971058042848}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368761344, u64=139971057970688}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368834208, u64=139971058043552}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368805344, u64=139971058014688}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368723736, u64=139971057933080}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368803760, u64=139971058013104}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368834032, u64=139971058043376}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368765568, u64=139971057974912}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368770848, u64=139971057980192}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368831568, u64=139971058040912}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368831392, u64=139971058040736}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368831216, u64=139971058040560}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368815904, u64=139971058025248}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368831040, u64=139971058040384}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368830864, u64=139971058040208}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368830688, u64=139971058040032}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368830512, u64=139971058039856}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368844592, u64=139971058053936}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368844416, u64=139971058053760}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368844240, u64=139971058053584}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368723384, u64=139971057932728}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368820656, u64=139971058030000}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368767856, u64=139971057977200}}, {EPOLLIN|EPOLLOUT|EPOLLERR|EPOLLHUP|0x2000,
{u32=2368844064, u64=139971058053408}}, {EPOLLOUT|EPOLLHUP, {u32=2368771200,
u64=139971057980544}}}, 128, 0) = 34 <0.000010>
1406320197.607291 26587 write(439, "hi", 2)   = -1 ECONNREFUSED (Connection
refused) <0.000006>
1406320197.607607 26587 epoll_ctl(4, EPOLL_CTL_DEL, 439, {0, {u32=4213032,
u64=4213032}}) = 0 <0.000006>
1406320197.607631 26587 close(439)            = 0 <0.000010>

Attachments:

  1. test.c (1874 bytes)
@dvyukov

This comment has been minimized.

Copy link
Member

commented Jul 26, 2014

Comment 1:

To make it clear, the effect of this bug is that connect returns OK when it should not,
but then the first read/write returns ECONNREFUSED, right?
Will waiting for only EPOLLIN after connect fix it?
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 26, 2014

Comment 2 by jason@eggnet.com:

No, connect still returns EINPROGRESS. But adding the descriptor to epoll prior to
calling connect causes an event for this descriptor to show up in epoll_wait being
executed in another thread.
That causes go to believe the descriptor is ready after calling connect later on. go
then calls getsockopt which clearly doesn't consider EINPROGRESS an error, thus
returning 0. go believes connect succeeded and moves on to to the write.
Only looking for EPOLLIN is against the man page. The proper flag to look for is
EPOLLOUT like go already does.
The problem is simply that go is polling a file descriptor prior to connect, which
appears to lead to undefined behavior.
@dvyukov

This comment has been minimized.

Copy link
Member

commented Jul 26, 2014

Comment 3:

I agree than we need to call epoll_ctl(ADD) after connect at least on linux. I don't
know about darwin/kqueue and solaris/ports.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 26, 2014

Comment 4 by jason@eggnet.com:

Agreed. Solving this for multiple platforms makes this a lot more complicated.
I also noticed one more problem with just triggering on EPOLLIN. The second epoll_wait
call that has this event also does not have EPOLLIN set. If go just looks for EPOLLIN,
it will probably just leave this fd open forever.
I guess the question is do you have the ability to test something on all of the
supported go target environments?
It's pretty hard to replicate the problem in go. On the other hand replicating it in C
should be straight forward.
A program could be written to spawn a thread in epoll_wait (or the kqueue, etc
equivalents on relevant platforms) to determine which operating systems exhibit the
problem in the first place. A program could also be written to test the behavior of
polling after connect, basically a portable version of the test.c program I uploaded.
I was digging around in libev, since it is a popular cross platform event library, for a
potential answer to this question.
The library is just a thin abstraction layer around various event handling mechanisms. I
guess that is why it is popular. In short, the library doesn't call connect. That is
something the user of the library is supposed to do.
But I think libev would be useful for simplifying the creation of a portable test.c
program since it covers all of the supported go platforms.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 27, 2014

Comment 5 by jason@eggnet.com:

The C tests I suggested may not be necessary.
I took a look at libuv, used by node.js, rust and others. Unlike libev, libuv has a
connect function.
uv__tcp_connect is defined in this file:
https://github.com/joyent/libuv/blob/9b4f2b84f10c96efa37910f324bc66e27aec3828/src/unix/tcp.c
line 127 is the connect call
line 149 adds the file descriptor to the poller.
This is platform agnostic behavior with the exception of windows.
Here is the code for windows https://github.com/joyent/libuv/blob/master/src/win/tcp.c
handle->func_connectex is called before REGISTER_HANDLE_REQ, which as far as I can tell,
means connect is called before the polling function on windows as well.
libuv appears to have unit tests for all of this.
@dvyukov

This comment has been minimized.

Copy link
Member

commented Jul 27, 2014

Comment 6:

I am sure that windows is fine. I propose to not touch it w/o indication of it being
broken.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 27, 2014

Comment 7 by jason@eggnet.com:

Agreed.
@mikioh

This comment has been minimized.

Copy link
Contributor

commented Jul 28, 2014

Comment 8:

thanks for the investigation and great help. here is a fix;
https://golang.org/cl/120820043/
though, not tested on dragonfly (async-connect enabled platform) yet.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 28, 2014

Comment 9 by jason@eggnet.com:

Thanks, this looks great.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 28, 2014

Comment 10 by jason@eggnet.com:

I've applied the patch and it seems to work. I can't reproduce any issues, thanks!
@gopherbot

This comment has been minimized.

Copy link
Author

commented Jul 28, 2014

Comment 11:

CL https://golang.org/cl/120820043 mentions this issue.
@mikioh

This comment has been minimized.

Copy link
Contributor

commented Jul 29, 2014

Comment 12:

Labels changed: added release-go1.3.1.

@mikioh

This comment has been minimized.

Copy link
Contributor

commented Jul 29, 2014

Comment 13:

This issue was closed by revision c0325f5.

Status changed to Fixed.

@rsc

This comment has been minimized.

Copy link
Contributor

commented Aug 11, 2014

Comment 14:

This is a big change and the CL description just says this "could be" a fix for
something.
I believe the code in 1.3 was the same as in 1.2, and somehow everything was fine in 1.2.
Is this urgent enough to warrant inclusion in 1.3.1? I don't see a clear argument for
that.
@rsc

This comment has been minimized.

Copy link
Contributor

commented Aug 11, 2014

Comment 15:

Issue #8276 has been merged into this issue.

@rsc

This comment has been minimized.

Copy link
Contributor

commented Aug 11, 2014

Comment 16:

The report in 8276 suggests that this did change in 1.3. The question is now just
whether it is urgent enough.
@rsc

This comment has been minimized.

Copy link
Contributor

commented Aug 11, 2014

Comment 17:

Issue #8276 has been merged into this issue.

@gopherbot

This comment has been minimized.

Copy link
Author

commented Aug 11, 2014

Comment 18 by jason@eggnet.com:

It is urgent for issue #8276 and applications like it, that infer something significant
from a successful connect. It could be a monitoring application that infers
connectivity, or maybe an http proxy that can fail over on connect but not write.
Of course, along with enough traffic/concurrency to trigger the problem.
@gopherbot

This comment has been minimized.

Copy link
Author

commented Aug 12, 2014

Comment 19 by martin.garton@ft.com:

Triggering the problem needs very little traffic or concurrency.  The following code 
shows the problem in milliseconds for me.
for {
    // replace 1.2.3.4 with a host that is down
    _, err := net.DialTimeout("tcp", "1.2.3.4:12345", time.Second)
    if err == nil {
        panic("Oops. We 'connected' to an non-existing host")
    }
}
(NB. If the host is listening, or returns a RST, or you get ICMP Destination unreachable
then this may not show the problem.)
If the proposed CL is believed risky, can we consider reverting the change that
originally broke this?:
https://code.google.com/p/go/source/detail?r=5f662f12d550
That would at least get us back to 1.2 functionality.
@rsc

This comment has been minimized.

Copy link
Contributor

commented Aug 12, 2014

Comment 20:

Approved for Go 1.3.1.
This code needs to stop changing. It seems to get rewritten on every release.
We can't work toward stability when things get rewritten so often.
If the rewrites can't stop on their own, we will have to introduce an explicit
freeze for large parts of package net.
Please try to wind them down.
@mikioh

This comment has been minimized.

Copy link
Contributor

commented Aug 13, 2014

Comment 21:

> Please try to wind them down.
yup.
> It seems to get rewritten on every release.
for the record, the bug was introduced in go1.1 by
https://code.google.com/p/go/source/detail?r=fed789ce8072 and i overlooked that nb "new
network poller can have spurious
readiness notifications" when i fixed the issue regarding async-connect feature on
dragonfly in go1.3. so pls blame me, ugh.
@adg

This comment has been minimized.

Copy link
Contributor

commented Aug 13, 2014

Comment 22:

This issue was closed by revision 073fc578434b.

@adg

This comment has been minimized.

Copy link
Contributor

commented Aug 13, 2014

Comment 23:

This has been applied to release-branch.go1.3.

@rsc rsc added this to the Go1.3.1 milestone Apr 14, 2015

@rsc rsc removed the release-go1.3.1 label Apr 14, 2015

adg added a commit that referenced this issue May 11, 2015

[release-branch.go1.3] net: prevent spurious on-connect events via ep…
…oll on linux

««« CL 120820043 / 06a4b59c1393
net: prevent spurious on-connect events via epoll on linux

On Linux, adding a socket descriptor to epoll instance before getting
the EINPROGRESS return value from connect system call could be a root
cause of spurious on-connect events.

See golang.org/issue/8276, golang.org/issue/8426 for further information.

All credit to Jason Eggleston <jason@eggnet.com>

Fixes #8276.
Fixes #8426.

LGTM=dvyukov
R=dvyukov, golang-codereviews, adg, dave, iant, alex.brainman
CC=golang-codereviews
https://golang.org/cl/120820043
»»»

TBR=r, rsc
CC=golang-codereviews
https://golang.org/cl/128110045

@golang golang locked and limited conversation to collaborators Jun 25, 2016

wheatman added a commit to wheatman/go-akaros that referenced this issue Jun 25, 2018

net: prevent spurious on-connect events via epoll on linux
On Linux, adding a socket descriptor to epoll instance before getting
the EINPROGRESS return value from connect system call could be a root
cause of spurious on-connect events.

See golang.org/issue/8276, golang.org/issue/8426 for further information.

All credit to Jason Eggleston <jason@eggnet.com>

Fixes golang#8276.
Fixes golang#8426.

LGTM=dvyukov
R=dvyukov, golang-codereviews, adg, dave, iant, alex.brainman
CC=golang-codereviews
https://golang.org/cl/120820043

wheatman added a commit to wheatman/go-akaros that referenced this issue Jul 9, 2018

net: prevent spurious on-connect events via epoll on linux
On Linux, adding a socket descriptor to epoll instance before getting
the EINPROGRESS return value from connect system call could be a root
cause of spurious on-connect events.

See golang.org/issue/8276, golang.org/issue/8426 for further information.

All credit to Jason Eggleston <jason@eggnet.com>

Fixes golang#8276.
Fixes golang#8426.

LGTM=dvyukov
R=dvyukov, golang-codereviews, adg, dave, iant, alex.brainman
CC=golang-codereviews
https://golang.org/cl/120820043

This issue was closed.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
You can’t perform that action at this time.