Bug #15386
[PATCH] io.c (rb_io_check_char_readable): do not io_fflush buffered sockets
Description
This is a rare corner-case probably nobody cares about. (because socket has IO#sync=false by default). Not critical to fix or change before 2.6 (I'm not sure if there's a compatibility risk). Would much appreciate an extra set of eyes to review; but it can wait for post-26. [I tried taking a break from C and ruby-core to work on a different project today; but hit this bug within an hour of doing that :<] io.c (rb_io_check_char_readable): do not io_fflush buffered sockets I enabled userspace buffering on sockets to reduce syscall overhead while avoiding non-portable TCP_CORK/TCP_NOPUSH. This deadlocked for me because I was using independent threads for reading and writing simultaneously on the same socket. I also experimented with making io_fflush optionally non-blocking, but that caused stream corruption with the reader thread doing some writes. https://80x24.org/spew/20181206104008.29153-1-e@80x24.org/raw Test script (hitting news://news.public-inbox.org/inbox.comp.version-control.git is fine) require 'socket' require 'uri' require 'io/nonblock' usage = "usage: #$0 news://news.public-inbox.org/inbox.comp.version-control.git" uri = ARGV.shift or abort usage uri = URI(uri) uri.port ||= 119 group = uri.path.sub(%r{\A/+}, '') # String#delete_prefix requires Ruby 2.5+ s = Socket.tcp(uri.host, uri.port) l = s.gets l =~ /\A2\d\d / or abort "bad greeting: #{l}" s.nonblock = true s.puts "GROUP #{group}" l = s.gets code, _, min, max = l.chomp!.split.map!(&:to_i) code == 211 or abort "bad GROUP response: #{l}" rdr = Thread.new do nres = 0 r = s.dup while l = r.gets l.start_with?('205 ') and break # cmd_quit l.start_with?('224 ') or abort "bad OVER response: #{l}" while l = r.gets if l == ".\r\n" nres += 1 break end end end nres end range = min..max s.sync = false range.each { |i| s.puts "XOVER #{i}" } puts "requests=#{range.size} #{Time.now}" s.puts "QUIT" s.flush puts "responses=#{rdr.value} #{Time.now}"
Files