Why doesn't rate limiting with trickle work with ruby?

Question

I'm attempting to limit the bandwidth available to a running ruby process to prevent it flooding our link and affecting other traffic. The process is downloading large files over HTTP.

I've found numerous people suggesting trickle as a convenient userland-only option (like https://unix.stackexchange.com/a/34123/160146).

I've tested it with an scp download and confirmed it limits the download rate:

trickle -s -d 50 scp [email protected] localfile.dat

I've also tested a python 2.7 script and confirmed it limits the download rate:

# cat test.py
import urllib

testfile = urllib.URLopener()
testfile.retrieve("http://example.com/bigfile.dat")
# trickle -s -d 50 python test.py

Finally, I've tested with a basic ruby script and trickle has no effect. The scripts maxes out my downstream bandwidth:

# cat test.rb 
require 'net/http'

Net::HTTP.get_response(URI.parse("http://example.com/bigfile.dat"))
# trickle -s -d 50 ruby test.rb

I've tested with ruby 2.1, 2.2 and 2.3 with no difference in behaviour.

I notice the trickle documentation says it only works with dynamically linked programs, but I'm fairly sure thats the case for me. I'm using the ruby provided by Debian in /usr/bin, and ldd reports linked libraries:

# ldd /usr/bin/ruby2.3 
        linux-vdso.so.1 (0x00007fffdcdc6000)
        libruby-2.3.so.2.3 => /usr/lib/x86_64-linux-gnu/libruby-2.3.so.2.3 (0x00007f277e11a000)
        libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f277defd000)
        libgmp.so.10 => /usr/lib/x86_64-linux-gnu/libgmp.so.10 (0x00007f277dc79000)
        libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f277da75000)
        libcrypt.so.1 => /lib/x86_64-linux-gnu/libcrypt.so.1 (0x00007f277d83e000)
        libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f277d538000)
        libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f277d194000)
        /lib64/ld-linux-x86-64.so.2 (0x0000558796396000)

Interestingly, trickle can rate limit my ruby process if I use the curb (libcurl bindings) gem:

# cat test.rb
require 'curb'
http = Curl.get("http://example.com/bigfile.dat"))
# trickle -s -d 50 ruby test.rb

Does that suggest that ruby's default Net::HTTP library uses sockets in a way that's incompatible with trickle?

I'm out of ideas. Is there a fixed reason why trickle doesn't work with ruby or is there something I can do to get them working together?

Answer 1

I believe I've tracked this down. It's arguably a fault in trickle that's occurred as a result of creeping featurism in the socket() library call.

The trickle code redefines socket(), as you know. At the top of its reimplementation in the file trickle-overload.c it has a check that includes type == SOCK_STREAM so that only connections of type SOCK_STREAM are eligible for processing.

The ruby code opens the socket with type = SOCK_STREAM | SOCK_CLOEXEC. The SOCK_CLOEXEC is a Linux extension (since 2.6.27) that avoids the use of a subsequent fcntl call to set the close-on-exec flag.

Unfortunately SOCK_STREAM is not the same as SOCK_STREAM | SOCK_CLOEXEC so the trickle processing will not take place until its code is fixed to allow for the extensions now permitted within socket's type field.

I know of no workaround other than to recompile trickle for your system with a modified socket type comparison.

(I've submitted a bug report to Debian Bugs since that's my preferred platform.)