RequestBuilder::send() can result in SIGPIPE

[PeterSP]

In a long-running process using hyper 0.6.14, inevitably after some hours of operation a SIGPIPE is received during hyper::net::HttpStream.Write::write calling net::tcp::TcpStream.Write::write calling send() from libc.

I'm under the impression this is an example of this sort of behavior, which could be fixed by catching the SIGPIPE.

Assuming I have the right idea, I would suggest either making sure not to send to closed sockets (which may not be easily possible), handling such SIGPIPEs in hyper, or adding documentation suggesting developers handle any SIGPIPEs which may come up while sending HTTP requests through hyper.

[seanmonstar]

Do you have any more specific info, such as OS, if using child processes, or calling in to Rust from another language, or anything?

Rust specifically turns off SIGPIPE behavior here: https://github.com/rust-lang/rust/blob/master/src/libstd/sys/unix/mod.rs#L76

So I'd think that if something is occurring that skips that, it should probably be fixed in Rust itself.

[PeterSP]

$ uname -a
Linux [hostname] 4.0.4-301.fc22.x86_64 #1 SMP Thu May 21 13:10:33 UTC 2015 x86_64 x86_64 x86_64 GNU/Linux

Using glibc 2.21

rustc version 1.3.0

Single process, run in gdb.

No other languages in the program (communicates over network with programs in other languages, but no other languages embedded in the program itself).

[seanmonstar]

Hm, some superficial Googling suggests that even with signal(SIGPIPE, SIG_IGN), gdb will still notice the signal and exit? Example http://ahlamnote.blogspot.com/2006/12/gdb-ignore-sigpipe.html

[PeterSP]

Bizarre. Hopefully that'll "fix" the "problem."

[seanmonstar]

I'm going to say that and SIGPIPE happening in Rust program that uses libstd is perhaps a bug in libstd.