software panacea

Tracking down a fd leak

2009-11-17T20:47:00.000-05:00

Once in a while, a system I worked on would generate a lot of these two errors at the same time.

IOError: [Errno 24] Too many open files

can't create socket

I know, it's "everything is a file" in unix so that explains "can't create socket".

I knew it was leaking file descriptors but I had to find the source. The corporate environment once again poses additional challenges. Corporate Policy prevents me from poking around on production servers.

Thankfully I have an awesome XML-RPC console built-in my application. It runs regular python code within the server process. The application and it's designers deserve far more credit than I gave them in a previous post. This feature is a god send.

>>> import os

>>> os.listdir('/proc/%s/fd' % os.getpid())

['0', '1', '2', '3', '4', '5', ... '889']

Uh oh! This isn't good. What are these files?

>>> data = os.fdopen(889).read()

>>> send_mail('mycorporateemail@email.com', 'Your file', 'Empty body', attachments={'data.bin': data})

After a few misguided attempts I open this file as a with Microsoft Word. It is an attachment an agent made when submitting a new client.

I quickly need to understand how attachments are generated in our framework. It was written with layers of WSGI which is a good thing because WSGI is simple to understand.

The application uses the standard python module cgi.FieldStorage to handle form attachments. The class documentation reads:

This version opens a temporary file for reading and writing,
and immediately deletes (unlinks) it. The trick (on Unix!) is
that the file can still be used, but it can't be opened by
another process, and it will automatically be deleted when it
is closed or when the current process terminates.

Closing the "file" contained in cgi.FieldStorage, does not release the file descriptor as stated. I have not determined if this is a bug in python or my application. However, I patched my application by adding a __del__ method to our request object. The object holds a reference to the file, and my __del__ will os.close the file's fd number. This is the only time I've ever had to implement __del__. I was thankful that the request object was not involved in a circular reference leak.

tomato router firmware is awesome

2009-11-14T10:11:00.000-05:00

I've tried a few different firmwares for my linksys wrt54g router and I finally settled on the tomato firmware. It's far better than others I've tried. Here's what sets it apart:

* The system is responsive and fast
* The web interface is well thought out and designed
* The setup and configuration processes work well and are full-featured
* It has nice graphs and reports about bandwidth usage built-in

I highly recommend it.

search and replace on many files in vim

2007-08-01T13:29:00.000-04:00

I'm posting this so I can catalog this tip for my own future reference,

argdo! %s/{%= \(.*\)\.label
%}<\/label>/{% label \1 %}/ge update

Compiling APUE source on OpenBSD

2006-12-19T19:43:00.000-05:00

I got the books source code from http://www.apuebook.com/ and tried to compile it under OpenBSD.

The first thing I tried was make the main Makefile include 'freebsd.mk', but some files still wouldn't compile.

As a result, I removed socket/recvfd2.c and socket/sendfd2.c from the build because OpenBSD does not support SO_PEERCRED. It provides getpeereid(2), which I will look up later.

Next, OpenBSD doesn't provide the POSIX.2001 pseudo-terminal functions grantpt(), ptyname(), posix_openpt(), unlockpt... a bit of googling brought up this short thread:
http://marc.theaimsgroup.com/?l=netbsd-tech-kern&m=108534411826666&w=2

After reading it, I gather that not providing those functions is a good thing: one less suid binary in the system. Better yet, the same functionality it still provided by OpenBSD, Quoting from same post:

OpenBSD implemented a different (better) way of doing this, by adding
/dev/ptm, and getting the master, slave fd's and the master and slave
names through an ioctl(3), doing the chown/chmod's in the process.

The posix way, requires a userland setuid program ``pt_chown'' to do
the nasty chown job, which is called from grantpt(3).