Strange interference between LSAPI, popen3 and MySQL

[Andreas Schwarz]

Hi,

I am experiencing a very strange problem in my Rails app. I use popen3
to execute an external program. Everything works fine in Mongrel, but
when I use the app with LSAPI, after popen3 all MySQL queries fail with
"Mysql::Error (MySQL server has gone away)". Once I comment out popen3
it works fine again.

Does anyone have the slightest idea about what goes wrong here?

Thanks
Andreas

 

[Andreas Schwarz]

The problem is easy to reproduce in any Rails app using LSAPI and MySQL.
Add the following to a controller action:

require 'open3'
Open3.popen3('/usr/bin/ls') {|stdout, stdin, stderr| stdin.read }

The stdin.read is important.

 

[ara.t.howard]

Hi,

I am experiencing a very strange problem in my Rails app. I use popen3
to execute an external program. Everything works fine in Mongrel, but
when I use the app with LSAPI, after popen3 all MySQL queries fail with
"Mysql::Error (MySQL server has gone away)". Once I comment out popen3
it works fine again.

Does anyone have the slightest idea about what goes wrong here?

guessing, but popen3 uses fork. forking with and open db handle is undefined
for almost any db lib - of course people ignore this. i did, and had all
sorts of issues. you can read about them here

http://www.linuxjournal.com/article/7922

i bet what happens is that the parent (rails) has setup an at_exit handler to
close the db on exit. when popen forks and does it's thing in the child, the
child's death kills the db handle (at_exit handlers being inherited - of
course).

this shows what i'm talking about


harp:~ > cat a.rb
at_exit{ puts 'close connection __A__' }
fork{ }
Process.wait

class Object
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end
module Kernel
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end

at_exit{ puts 'close connection __B__' }
fork{ }
Process.wait


harp:~ > ruby a.rb
close connection __A__


i'd try putting this


class Object
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end
module Kernel
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end

in environment.rb or something. another idea, i'm grasping, is some sort of
buffer flushing thing.

you could try


4242.times do |fd|
begin
IO.for_fd(fd).flush
rescue Errno::EBADF
end
end

right before the popen. wild ass guess though.

good luck.

-a

 

[Andreas Schwarz]

I am experiencing a very strange problem in my Rails app. I use popen3

guessing, but popen3 uses fork. forking with and open db handle is undefined
for almost any db lib - of course people ignore this. i did, and had all
sorts of issues. you can read about them here

OK, good to know. I can close the MySQL connection for the popen3 call
and reopen it after it to avoid problems.

i bet what happens is that the parent (rails) has setup an at_exit handler to
close the db on exit.

That makes sense, although I couldn't find any reference to at_exit in
Rails or ruby-lsapi.

i'd try putting this


class Object
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end
module Kernel
alias_method '__fork__', 'fork'
def fork *a, &b
at_exit{ exit! }
__fork__ *a, &b
end
end

in environment.rb or something.

This fixes the MySQL problem (as does manually
disconnecting/reconnecting), but now I have another problem: popen3 just
doesn't do anything when used within LSAPI. The program is never
executed. Any ideas?

Who would have thought that simply calling an external program could be
so difficult...

Thanks
Andreas

 

[ara.t.howard]

and reopen it after it to avoid problems.


That makes sense, although I couldn't find any reference to at_exit in
Rails or ruby-lsapi.

huh - maybe it's the actuall c mysql driver.

This fixes the MySQL problem (as does manually
disconnecting/reconnecting), but now I have another problem: popen3 just
doesn't do anything when used within LSAPI. The program is never
executed. Any ideas?

hmmm. no. must be related though. i'm staring to thinks it's lsapi which has
setup all the funky at_exit handlers, and that this is causing you grief...

Who would have thought that simply calling an external program could be
so difficult...

me ;-)

ultimately, in ruby queue, i had to setup a drb process before setting up db
connections. then i used that drb process to do all forking/waiting for me.

here's another idea, try systemu, it does not use fork at all

gem install systemu

http://rubyforge.org/frs/?group_id=1024&release_id=7721
http://codeforpeople.com/lib/ruby/systemu/systemu-1.0.0/README

unless you need realtime control over stdin/stdout/stderr it'll do the trick.

regards.


-a

 

[Andreas Schwarz]

hmmm. no. must be related though. i'm staring to thinks it's lsapi
which has
setup all the funky at_exit handlers, and that this is causing you
grief...

I grepped for at_exit in the LSAPI source code, but didn't find
anything. Maybe it's the stdout/stderr/stdin redirection stuff...

me ;-)

ultimately, in ruby queue, i had to setup a drb process before setting
up db
connections. then i used that drb process to do all forking/waiting for
me.

here's another idea, try systemu, it does not use fork at all

gem install systemu

http://rubyforge.org/frs/?group_id=1024&release_id=7721
http://codeforpeople.com/lib/ruby/systemu/systemu-1.0.0/README

It seems to use popen, isn't this based on fork, too?

The reason I want to use popen3 is not that I need stdin/stderr, but
that there is a clean way to pass arguments without having to build a
command line (potential escaping problems etc.).

 

[Ezra Zygmuntowicz]

Hi~

I grepped for at_exit in the LSAPI source code, but didn't find
anything. Maybe it's the stdout/stderr/stdin redirection stuff...


It seems to use popen, isn't this based on fork, too?

The reason I want to use popen3 is not that I need stdin/stderr, but
that there is a clean way to pass arguments without having to build a
command line (potential escaping problems etc.).

LSAPI does its own preforking of Rails processes and disconnects and
reconnects the database handles after every fork. The way they manage
ruby processes causes weird issues like the one you are seeing. The
manager code that does the fork management is closed source so it's
hard to know exactly what is going on in there.

Cheers-
-- Ezra Zygmuntowicz
-- Lead Rails Evangelist
-- (e-mail address removed)
-- Engine Yard, Serious Rails Hosting
-- (866) 518-YARD (9273)

 

[Ezra Zygmuntowicz]

It's all under a BSD-looking license:
gem install ruby-lsapi

jeremy

You're right, my bad. I didn't realize they distributed the C code
that does the management as well.


-- Ezra Zygmuntowicz
-- Lead Rails Evangelist
-- (e-mail address removed)
-- Engine Yard, Serious Rails Hosting
-- (866) 518-YARD (9273)

 

[George Wang]

OK, we figured out what is wrong. "reopen" method need to be defined for
"LSAPI" class since popen3 calls "reopen". LSAPI new release will
address this.

However, the root cause of broken MySQL connection is not LSAPI's fault,
actually a bug in ruby's "open3.rb", the fix is like:

def popen3(*cmd)
pw = IO:ipe # pipe[0] for read, pipe[1] for write
pr = IO:ipe
pe = IO:ipe

pid = fork{
# child
fork{
# grandchild
+ begin

pw[1].close
STDIN.reopen(pw[0])
pw[0].close

pr[0].close
STDOUT.reopen(pr[1])
pr[1].close

pe[0].close
STDERR.reopen(pe[1])
pe[1].close

exec(*cmd)
+ ensure
+ exit!(0)
+ end
}
exit!(0)
}

pw[0].close
pr[1].close
pe[1].close
Process.waitpid(pid)
pi = [pw[1], pr[0], pe[0]]
pw[1].sync = true
if defined? yield
begin
return yield(*pi)
ensure
pi.each{|p| p.close unless p.closed?}
end
end
pi
end
module_function open3
end

Whenever an error raise in the grandchild process, this process will not
be terminated properly, and DB connection was closed in this process,
thus the DB connection error in the main process. It happens to both
mongrel and LSAPI.

With above fix, DB connection is in good shape even without the LSAPI
fix, just the command was not getting executed.

Best Regards,
George Wang
http://www.litespeedtech.com/