Exceed reader limits

[nbaum]

Possible fix for #10282, #10655.

This introduces readbytes(cmd::AsyncStream). Is it bad form to introduce new methods in a bugfix?

No tests. Sorry.

[nbaum]

The tests fail at line 263 of repl.jl, output = readall(master). Probably this is related to master being a TTY.

Annoyingly, if I write to STDOUT just before the readall, the test passes...

[nbaum]

I believe this has exposed something that was going wrong anyway.

The libuv read callback is being invoked with a negative nread: -5 for me, perhaps for everyone. It looks like it happens instead of an invocation with nread == UV_EOF.

The callback notifies readnotify waiters that an error occurred, but doesn't notify_error closenotify waiters. It forceably closes the stream, which then causes the close callback to be invoked: then the closenotify waiters are notified, but not told there was an error.

I don't know why the read callback is being invoked with a negative nread.

[nbaum]

The error is UV_EIO. I believe it relates to the interaction between the TTY master and its slaves.

This SO question says that closing all the slaves on a TTY causes reading from the master to fail with EIO.

It suggests keeping a "spare" slave open, but if we do that readall never returns because the master never closes. I've adjusted the test to use readavailable, and then close the slave descriptor, which works.

The updated PR also changes the callback handler to notify close-waiters. Should that actually be in a separate PR, or is it okay in this one?

Aside: It looks like writing to stdout hides the problem because by the time the write has finished, the subprocess has already exited and closed the slave so readbytes doesn't ever wait for the readnotify condition and it never sees the error. The error state isn't stored anywhere permanent, so it looks like we can never know it happened. That feels like a bug.

[nbaum]

I can't make make test fail on my x86_64 system here. I don't know what to do about this.

[vtjnash]

i'm looking into possible fixes for this