Recently there were discussions and blog posts on Docker for Pharo and Gemstone/S. This is my report after spending an afternoon on the subject.
First, some links:
This blog is implemented in Pharo and is the natural choice for my Docker example application. I already have a Smalltalk snippet to load this blog's code and its dependencies into a pristine Pharo image, so I'll be using that. Also, as a matter of course, I build the Pharo VM from source, and my VM installation also contains self-built shared libraries like libsqlite.so and libshacrypt.so.
Outside of Docker, prepare a custom Pharo image:
gofaro is a simple shell script which purpose is to make sure the Pharo VM loads my custom shared libraries, co-located with the standard VM files, at run time:
loadSCMS1.st looks like this:
Before describing my Dockerfile, here are my conventions for inside the Docker container:
Starting with Ubuntu 18.04, install libfreetype6. The other lines are copied from Torsten's tutorial.
Next, install the Pharo VM.
Now copy over the prepared Pharo image.
Finally, set the Docker container running. Here we create a UID/GID pair to run the application. Said UID owns the mutable Pharo files in /pkg/image and also the /pkg/image directory itself, in case the application needs to create other files such as SQLite databases.
runSCMS1.st runs the blog application. In my current non-Dockerized installation, the runSCMS1.st-equivalent snippet is in a workspace; for Docker, to become DevOps/agile/CI/CD buzzwords-compliant, this snippet is run from the command line. This is one Dockerization adaptation I had to make to my application.
Now we build the Docker image.
The Docker image has been created, but it is not ready to run yet, because the web content is not in the image. I'll put the content in a Docker volume. Below, the first -v mounts my host's content directory into /tmp/webcontent in the container; the second -v mounts the volume smdw-content into /pkg/cms in the container; I'm running the busybox image to get a shell prompt; and within the container I copy the web content from the source to the destination.
Finally, run the Docker image, taking care to mount the volume smdw-content, now with this blog's content:
Verified with a web browser. This works on my computer. :-)
According to its documentation, on Unix, the Pharo VM's SSL plugin, libSqueakSSL.so, links into OpenSSL libraries dynamically. On my 64bit Ubuntu Trusty machine, OpenSSL is provided by the libssl1.0.0:i386 package.
(By the way, the SSH2 plugin libssh2.so.1.0.1 requires libcrypto too.)
According to packages.ubuntu.com, Trusty's libssl1.0.0 is built from openssl_1.0.1f.orig.tar.gz plus successive upstream patches.
From the OpenBSD developers, LibreSSL is "a version of the TLS/crypto stack forked from OpenSSL in 2014, with goals of modernizing the codebase, improving security, and applying best practice development processes." LibreSSL also comes with libtls, "a new TLS library, designed to make it easier to write foolproof applications".
Let's see how we go about linking libSqueakSSL.so with LibreSSL.
First, download and unpack LibreSSL. Modify the configure script at lines 2287 and 2289 so that LIBCRYPTO_VERSION and LIBSSL_VERSION both say 1:0:0 instead of 35:0:0. Then build LibreSSL:
I'm building on a 64bit OS, hence "-m32". Without "--disable-asm", the build fails. To get the assembler version, which is recommended for serious usage, either set up a 32bit build environment or muck around with autoconf/configure. I suspect the former is easier. :-)
The output files are $SRC/crypto/.libs/libcrypto.so.1.0.0 and $SRC/ssl/.libs/libssl.so.1.0.0. The shared object files have the "1.0.0" suffix because I modified configure above. Alternatively, I could've played around with autoconf, or built the shared objects with the "35.0.0" suffix and sym/hard-link them for the "1.0.0" versions. TIMTOWTDI.
Next, remove the OpenSSL package:
Finally, put the LibreSSL shared object files into the right place. Where this right place is depends on your environment. TIMTOWTDI. I choose to put them in the Pharo VM directory with its other plugins, and arrange to start Pharo with LD_LIBRARY_PATH set appropriately. Going by the output of ldd again, the following is required:
Launch the Pharo 4.0 image and run the Zodiac tests. All tests should pass. Well, except testGetPharoVersion, which looks for a file that apparently no longer exists.
Incidentally, Squeak 5.0-All-in-One's SSL plugin appears to have linked its crypto/SSL libraries in statically, so the only way to upgrade is to build a new plugin.
I run Debian 6 64bit on one of my servers. To run the 32bit Pharo/Cog VM, 32bit libraries are needed and are installed thusly:
# apt-get install ia32-libs
Trying out the newly released Ubuntu 14.04, I learn that ia32-libs is considered a hack and was deprecated in 2013 versions of Ubuntu. I guess the idea is to encourage people to know what they require more precisely.
To get the current Pharo-VM-linux-stable.zip to work on Ubuntu 14.04 64bit, install the following packages:
# apt-get lib32gcc1 # apt-get libssl0.9.8:i386
On libssl, that is, OpenSSL, Ubuntu also ships 1.0.1f-1ubuntu2, which patches the Heartbleed bug, so you may install that instead.
In deployment war stories, I wrote that Pharo's graphical environment is managed by MorphicUIManager, but when invoked -nodisplay, Pharo's graphical environment is managed by NonInteractiveUIManager.
I prefer running my servers without X as far as possible, so I want to have RFBServer running within the Pharo image, to be able to VNC into it.
As it turns out, modifying one method, MorphicUIManager>>onSnapshot:, gets me what I want. Here's what the method looks like in Pharo 1.4, timestamped CamilloBruni 2/13/2012 23:22, (minus comments because I don't like the way the Smalltalk syntax highlighting JS library I'm using is rendering them):
I just needed to add tests for RFBServer, so now the method becomes like this:
And now VNC into the Pharo image works, yay!
I used to run Squeak and subsequently Pharo on FreeBSD. I had the Smalltalk RFB server running within the image, and connected to the image via COTVNC (Chicken of the VNC), an OSX client. I ran Squeak/Pharo in a FreeBSD jail, and doing it this way kept the jail minimalist and tidy. No X stuff was installed.
Then I ran into some difficulty with FFI on CogVM on FreeBSD. (That's for another post.) And I was running FreeBSD 7, which EOLed. So I switched to Linux. Chose Debian 6 after trying out several distros. Similar setup as on FreeBSD: Minimalist LXC container, CogVM, RFB server in the same Pharo image that ran on FreeBSD.
Interestingly, neither COTVNC nor several other VNC clients (both within X on Linux and from OSX, including Java and native ones) worked in this setup: As far as I determined, the VNC connections were all failing because of incompatible encodings, and I tested quite a few combinations.
I wasn't interested in debugging too much, so I did the next simplest thing: Installed an Xvnc environment. So, instead of running an RFB server within Pharo, I had Pharo running within an X environment that was also a VNC server. And then all the VNC clients I was testing with worked. Not as tidy as I wanted, because now the LXC container had a whole bunch of X cruft, but hey, it was working.
But I really prefer FreeBSD. So after getting my production Linux server up and running, I went to investigate my FreeBSD CogVM FFI issue. And it turned out to be a process error on my part. I think. But that's for another post.
A note on my (fairly typical) setup: My development machine is an MBP. My development and pre-production FreeBSD and Linux servers all exist as VirtualBox or VMWare virtual machines on my MBP. The development servers have X, while the pre-production servers are meant to provide the canonical configuration for the production servers, so I run them without X, as much as possible. On FreeBSD, I have never had to run X in production. On Linux, I have the aforementioned Xvnc setup.
Anyhow, CogVM FFI FreeBSD was working again, in my development server running X. Next I tested it in my X-less pre-production server. When I invoked CogVM with "-headless", it said "cannot locate vm-display-X11" or words to that effect, even though said dynamic library was right there together with vm-display-null. I reckon that vm-display-X11 failed to locate an X environment, hence failed to load, hence the error message. Meaning, instead of -headless, I had to run -nodisplay.
So, CogVM ran -nodisplay, and COTVNC connected to it fine. Opened a workspace, typed some FFI-using code, selected the text with the mouse, pressed ESC. What's this? CogVM or Pharo crashed! PharoDebugLog recorded "SubclassResponsibility: NonInteractiveUIManager had the subclass responsibility to implement #newMenuIn:for:".
Hmmm. This didn't happen in Pharo 1.0 and 1.1, IIRC. Seems Pharo's random refactorings (heh, just kidding) changed how -nodisplay was managed between those versions and 1.4, the version I was testing.
After browsing a bit, looks like Pharo's graphical environment is managed by MorphicUIManager, but when invoked -nodisplay, even with a running RFB server, Pharo's graphical environment is managed by NonInteractiveUIManager.
One possibility to get this working is to install a minimal X environment, to run CogVM+Pharo -headless (with MorphicUIManager) instead of -nodisplay (with NonInteractiveUIManager).
But that isn't the simplest thing that could possibly work. What is? Copy & paste, of course. Specifically, I copied/pasted MorphicUIManager>>newMenuIn:for: into NonInteractiveUIManager. Ok, now pressing ESC brings up the context menu, "Do it" does, and output in the Transcript shows the expected FFI-related output.
Next I run my application's startup code in a workspace, one step of which is sending the "explore" message to the application's main instance, thus giving me a handle to said instance should I need to manipulate it, which is kinda the point of having VNC access into the image in the first place.
What's this? Crashed again! PharoDebugLog informs me "MessageNotUnderstood: NonInteractiveUIManager>>explorer:for:withLabel:" and "An attempt to use interactive tools detected, while in non-interactive mode".
Not so simple then...
(To be continued.)