Just For Fun: The Story of Linus Torvalds

For the last couple of weeks, my bedside reading has been this half-biography, half-autobiography on Linus Torvalds. I have to say, however, that the book is like two books mixed into one. Chapters alternate between Linus talking about his life and about big moments in Linux’s history to David Diamond describing modern-day Linus with a kind of forced wonder. Truthfully, Diamond comes off as a sycophant who could care less about Linus’s flaws and positive characteristics, and cares more about molding some kind of “image” of Linus as containing a humility and genius simultaneously. Near the end, I started only skimming the chapters not written by Linus. Diamond’s really not a good writer, either. (Sorry Dave.)

Truthfully, the book kind of pops the lid off Linux and makes you understand it as much less glamorous than say Wired Magazine described it to the public. Linus really just talks about not having a social life, sitting in his room with curtains covering his window, coding all day. Not exactly the ideal role model, I think. Don’t get me wrong, I love the Linux kernel (as much as one can love imperfect software), and Linus made a great contribution toward keeping the UNIX world and UNIX principles alive, but it’s just that I like to think of open source developers as something other than the stereotypical, introverted geek. In fact, much of Linus’s chapters is devoted to his apprehension about giving a public talk about Linux. When I think about the fact that I’ve given three or four of them to date, and enjoy it more every time, I see how different I am from this kind of stereotypical geek.

It also kind of made me dislike Linus. When I saw Revolution OS (a DVD on the rise of open source), the movie kind of endeared me to Linus’s practical nature as opposed to Richard Stallman’s religious idealism. I like idealism, but Stallman is really religious about it. And he’s bitter. Linus, on the other hand, has that great Northern European, “I’m just gonna go with the flow” attitude.

But this book made me realize that Linus is religious is his own sort of way. Included in the book is Linus’s flame war with Andy Tanenbaum on monolithic versus microkernel designs. Truthfully, I’ve studied operating systems and I’m not even sure which design is best, and Linus makes a decent argument of why microkernels end up being just as complex, or more complex than monolithic ones. But what I didn’t like is that in the flamefest, Tanenbaum said that deficiencies in MINIX were due to it being a hobby, and that he had duties as a professor. Linus responded, “Re 2: your job is being a professor and researcher: That’s one hell of a good excuse for some of the brain-damages of minix. I can only hope (and assume) that Amoeba [Tanenbaum’s future OS project] doesn’t suck like minix does.”

This just shows me that Linus really is an asshole sometimes. He states this outright in his book. So now, truthfully, I may like the open source movement, but I think I “at least dislike” two of its most major players (Torvalds and Stallman).

Finally, I think a clip from Tanenbaum’s website points out a nice principal in OS design:

Also, Linus and I are not “enemies” or anything like that. I met him once and he seemed like a nice friendly, smart guy. My only regret is that he didn’t develop Linux based on the microkernel technology of MINIX. With all the security problems Windows has now, it is increasingly obvious to everyone that tiny microkernels, like that of MINIX, are a better base for operating systems than huge monolithic systems. Linux has been the victim of fewer attacks than Windows because (1) it actually is more secure, but also (2) most attackers think hitting Windows offers a bigger bang for the buck so Windows simply gets attacked more. As I did 20 years ago, I still fervently believe that the only way to make software secure, reliable, and fast is to make it small. Fight Features.

I agree. But does a microkernel design actually reduce the overall size of the operating system, or does it just reduce the size of whatever you consider to be the “microkernel”? That is, just because a file system is implemented as a file system daemon talking to a driver subsystem through message passing doesn’t necessarily mean the file system, or driver subsystem, are secure. Insecurity could exist even at the boundaries, no? Not to mention instability.

I think Linus and Tanenbaum have to agree that this debate isn’t an open and shut case. The best kernel is probably one that mixes modularity, a strong kernel/userspace boundary, and some of the fancier features of a microkernel approach, while not sacrificing elegance of design or performance.

Keeping applications open

Just an interesting post I made to OSNews in response to someone saying that IE “starts quickly” while Firefox “takes forever.”

Just to clear up, the only reason IE starts faster in Windows is because IE is technically “always running.” The only thing that has to “start” is creating a Window with an “IE control” in it.

I get the same behavior on Linux by running galeon -s when my X session starts. This runs galeon in “server mode,” which means it’s always in memory, and when I run Galeon (on my laptop, I press ALT+F1 to run my browser), it starts in < a half-second. If Firefox had a similar mode, it could offer you the same thing. As for OpenOffice.org, it's true that the start time is relatively slow. I'm sure they'll get around to optimizing it. Personally, I think the obsession people have with start times on Linux and Windows machines is due to a basic design flaw with most Window managers. Applications should really only start up once; if you start an application multiple times in a day, you're essentially performing redundant computation. The program can sit in memory and if it really is not used in awhile, it will get paged out anyway due to our modern Virtual Memory implementations. In OS X, for example, you can get the same effect as "galeon -s" or IE's "preloading" simply by not quitting an application after all its windows are closed. This leaves the application running, and when you open a new window it will be nearly instantaneous. (Strangely enough, many old Windows/Linux freaks are sometimes "annoyed" by this aspect of OS X, since in the Linux/Windows world up to now, closing all windows of an application is equivalent to closing the application itself).

GNU ddrescue and dd_rescue and dd_rhelp, what the?

Wow. I hate when shit like this happens.

Apparently there are three tools out there to help with the same thing. First, there’s dd_rescue, the tool I was using earlier (which ships with Ubuntu in a debian package called… ddrescue). Then, there’s dd_rhelp, a shell script which is a frontend to ddrescue and which implements a rough algorithm to minimize the amount of time waiting on bad block reads.

Then, there’s GNU ddrescue, which is a C++ implementation of dd_rescue plus dd_rhelp.

I only just realized this and so now I’ve compiled a version of GNU ddrescue to pick up my recovery effort. It’ll probably help with one of the partitions that seems particularly messed up.

So far the nice thing about GNU ddrescue is that it seems faster, and more responsive. Plus, it has a real logging feature, such that if you enable it and then CTRL+C the app, you can restart it and it’ll automatically pick up where it left off.

UPDATE: wow, good thing I switched. GNU ddrescue is significantly faster just in terms of raw I/O performance. I jumped from 4GB of this partition being rescued (which took 30 minutes with dd_rescue) to 6GB in the last ten minutes. It seems at least 3x faster. I also like that the GNU info page describes the algorithmic approach in-depth.

Fried hard disk ruins weekend

So, one of my employers ended up with a fried hard disk, for the second time in a row. The main reason is that the PC this HD is contained in sits in a corner with little-to-no airflow.

In order to recover the drive, I am actually taking a different approach from my last recovery effort, mainly by necessity. This disk is seriously damaged–lots of bad sectors, and its partitions are not readable by any NTFS driver, be it Microsoft’s or the open source one. This makes simply using the wonderful R-Studio tool I used last time currently impossible, due to the fact that it won’t even see the drive properly within Windows, and will hang all over the place.

Indeed, what I needed to do is drop down a layer of abstraction: away from filesystems, and into blocks and sectors. Unfortunately, in the Windows world this drop down is difficult, so I had to use my Linux laptop to make this jump.

I found a wonderful tool to help me out called dd_rescue, which is basically a dd with the added features of continuing on error, allowing one to specify a starting position in the in/out files, and the ability to run a copy in reverse. These features allow one to really work around bad sectors and even damaged disk hardware to get as much data as possible out.

Unfortunately, the use of this tool was encumbered by my laptop’s relatively simple bus design. Apparently, if I stuck two devices on my USB bus (like two HDs I was using for this process), the bus would slow to a crawl, and the copy would move along at an unbearble 100kB/sec. I tried utilizing firewire and USB together, but got only marginal improvements. What befuddles me is that in the end, the fastest combination I could come up with is reading from the Firewire enclosure with my laptop and writing to the firewire enclosure of my desktop across the LAN utilizing Samba. Very strange indeed. Now my performance is more like 6MB/sec, factoring in all the breaks dd_rescue takes when it encounters errors. I have 6GB of the more critical partition written, but it’ll probably take a couple hours to have a big enough chunk that I can test R-Studio’s recovery of it.

The only reason I’m even writing about this is because I find it hilarious how many layers of abstraction I am breaking through to do a relatively low-level operation. Think about it:

  1. My broken IDE drive is converted to Firewire by a Firewire-IDE bridge.
  2. My Firewire PCMCIA adapter is allowing my notebook to take in that connection.
  3. The Linux kernel is allowing firewire to be accessed via various ieee1394 ohci drivers.
  4. The Linux kernel is abstracting the firewire disk as a SCSI disk, using emulation.
  5. The SCSI disk is being read by dd_rescue and written to a file, which exists in the path /mnt/smb/image/sdb5.
  6. That path seems local, but is actually a mount point. That mount point seems physical but is actually handled by a Samba driver.
  7. The writes by dd_rescue to that image file are being sent through the kernel’s TCP/IP stack, and flying through my switch, and being accepted by Windows XP’s network stack.
  8. Windows XP is writing that data to an NTFS drive, which is itself connected by a Firewire-IDE bridge (and therefore all the above steps’ equivalents for Windows apply).

I am surprised with that many layers, that this copy is even working. I really should have just taken a machine apart and connected these drives directly by IDE, to save myself a few layers.

N-way parallel mail retrieval with getmail and bash

I wrote a pretty sweet script tonight. It parallelizes the getmail retrieval process, while still printing prefixes so I know which accounts download which messages. This means that instead of my mail fetching process taking sum(i1,…,in), where i is the length of time for a given mail retrieval, my fetching process now takes max(i1,…,in).


#!/bin/sh
GETMAIL='python2.3 -Wignore /usr/bin/getmail'
unwanted() {
  grep -E -v '(Copyright|getmail|Simple)';
}

echo "N-WAY GETMAIL RETRIEVER SCRIPT:"
$GETMAIL \
  --rcfile=/etc/getmail/account1 \
  2>&1 | sed -e \
  's/.*/account1................: &/g' \
  | unwanted &

$GETMAIL \
  --rcfile=/etc/getmail/account2 \
  2>&1 | sed -e \
  's/.*/account2................: &/g' \
  | unwanted &

...

$GETMAIL \
  --rcfile=/etc/getmail/accountN \
  2>&1 | sed -e \
  's/.*/accountN................: &/g' \
  | unwanted &

wait

PIDA: Python Integrated Development Application

PIDA 0.2.2 was released recently. This is truly a novel development in the Python/OSS world. What PIDA provides is a nice plugin system and the “makings” of an IDE. So, in a nice IDE you have a class browser, an integrated debugger, a profiler, maybe even a RAD-like GUI builder, an interpreter console, etc. The one piece that tends to be most controversial in every IDE is the text editor. This is one-part because UNIX people are really crazy about their text editors, but two-parts because text editors are very important programmer tools, and no one wants to learn a different text editor for every language one uses.

vim happens to be awesome for C programming, which is probably why a lot of UNIX hackers use it. But for Python, more advanced support would be nice. PIDA can run and connect to a vim server instance in order to allow you to have an “add-on IDE” for vim.

But even more interesting to me is the culebra plugin, which provides a code-completion-savvy GtkSourceView inheritor, which has the initial support for fancy Intellisense-like features.

I’ve already spoken to the developers of PIDA, and they said they would very much be interested in seeing Python Intellisense features brought to VIM. When I started thinking about different approaches to doing this, I realized that the whole OSS community could benefit from a general Python module that enhances the Python introspection features (and perhaps combines them with source code parsing) to make available nice productivity-enhancing features. I was thinking of calling this beast “Pyductivity.”

More on that later. For now, check out PIDA.

Exa: a new architecture for Xorg

This is exciting news. A Trolltech developer has modified KAA, the acceleration architecture used in Keith Packard’s experimental “kdrive” Xserver, to work with the traditional Xorg tree. He announced this new development in an e-mail that makes it clear it is extremely easy to get drivers to use Exa to gain Apple/Windows-like graphics performance.

I know that the unichrome project generally doesn’t bother itself with these very desktop-oriented features (their focuses are more on MediaPCs, etc.), but I think this may be an excellent way for me to begin hacking the new modular Xorg tree I mentioned last time. If I added Exa support to the unichrome driver, would that mean transparency and full-on graphics acceleration for my X desktop, what I’ve long been waiting for? We’ll see.