openat(AT_FDCWD, "/tmp/fifo", O_RDONLY|O_NONBLOCK|O_CLOEXEC) = 3
read(3, "", 4096)                        = 0

The above strace log show that read(2) not return error when write end of the pipe has not been opened.

I tried to reproduce your comment by copygin my play.golang.org example above into bug.go, commenting out the demo(false) line in main, then go build bug.go and strace ./bug > /dev/null.

I see:

write(1, "\nRace\n", 6)                 = 6
mknodat(AT_FDCWD, "/tmp/fifo", S_IFIFO|0600) = -1 EEXIST (File exists)
openat(AT_FDCWD, "/tmp/fifo", O_RDONLY|O_NONBLOCK|O_CLOEXEC) = 3
epoll_create1(EPOLL_CLOEXEC)            = 4
pipe2([5, 6], O_NONBLOCK|O_CLOEXEC)     = 0
epoll_ctl(4, EPOLL_CTL_ADD, 5, {EPOLLIN, {u32=5884208, u64=5884208}}) = 0
epoll_ctl(4, EPOLL_CTL_ADD, 3, {EPOLLIN|EPOLLOUT|EPOLLRDHUP|EPOLLET, {u32=3547467400, u64=139637229227656}}) = 0
fcntl(3, F_GETFL)                       = 0x8800 (flags O_RDONLY|O_NONBLOCK|O_LARGEFILE)
fcntl(3, F_SETFL, O_RDONLY|O_NONBLOCK|O_LARGEFILE) = 0
futex(0xc000080148, FUTEX_WAKE_PRIVATE, 1) = 1
read(3, "", 4096)                       = 0
write(1, "n: 0, err: EOF\n", 15)        = 15
write(1, "b: []\n", 6)                  = 6
futex(0xc000042548, FUTEX_WAKE_PRIVATE, 1) = 1
epoll_ctl(4, EPOLL_CTL_DEL, 3, 0xc0000a1dac) = 0
close(3)                                = 0
futex(0x56d9e8, FUTEX_WAIT_PRIVATE, 0, NULL

at the point it hangs on wg2.Wait() as expected. But what I see differently to you is that there's several epoll calls in between the openat and the read. I don't understand what is happening here but I can guess it is because of the SetReadDeadline and the implementation of file IO in Go for Linux being based on epoll. I can make a wild guess that in an epoll-triggered read, there must be data waiting on the fd, rather than my earlier assumption of calling read on a not-ready-yet fd which would trigger EAGAIN.

But the result is the same -- if the read did not error (strace shows it didn't return -1) then the fd is not closed, and thus the higher level implementation should not be returning io.EOF in this case.

CC @ianlancetaylor, @bradfitz.

What sort of help do you need? :-) Would you like a change that implements what I suggested, or are you looking for an expert third party to approve of the proposed change first?

A change that implements the suggestion, along with a test, would be fine. Thanks.

I've drafted a test based on my playground code above, and so far have managed to figure out that this is going to be harder than I thought.

I'm stuck again, confusing myself in the depths of internal/poll and now runtime/netpoll.go

I'm reminded that golang/go/20280 is the reason I'm exploring this path. I have a fifo that I want to read from until either EOF or a context is cancelled. In order to notice that the context is cancelled, I use SetReadDeadline on the fifo, hence the strange construct in the playground example.

The earlier statements about expecting EAGAIN on read was incorrect, because Go runtime is using epoll in the runtime, so we should expect to see EPOLLHUP on the reader while the write end is still closed. I think the EPOLLHUP gets eaten in netpoll_epoll.go:164 and the intent of "the writer isn't there" never gets passed up, but I haven't verified yet.

(For the benefit of the tape, the best way to debug this that I've found is after having built a local go fresh from Git (~/src/go/go for me) use the new install's go to install delve (~/src/go/go/bin/go install github...delve@latest), and then cd into ~/src/go/go/src/os and run PATH=~/src/go/go/bin:$PATH GOROOT=~/src/go/go dlv test to run the current test with the correct toolchain and be able to recompile the test as you modify it to explore the code.)

Meanwhile I've noticed #20280 (comment) which I think is a better approach than setting a read deadline every time just to test for cancellation, so while I think there's still a bug here, I think I am no longer dependent on it.

Could you open fifo in blocking mode, that should avoid the race?

The problem I have is nearly identical to #20280, where one does not want to open in blocking mode because one wants a context cancellation to interrupt the read. The suggestions from that bug report have indeed helped -- the magic is knowing you don't need to set a read deadline on the read before you start the read.