OK, so here's a question. I've always wanted to spend some time reading through the Linux kernel - I really enjoyed operating systems when I studied it (and reading through MINIX src) but I don't really know where/how to start with Linux.
I mean, I guess I can just dive in, and maybe that's the best approach, but is there a strategy anyone would recommend for reading the Linux source in terms of it making sense as a combined unit of code (as opposed to a collection of algorithms, if that makes sense)?
I recently did exactly what you describe. My first contribution to the Linux kernel actually shipped with this release. If I recall correctly it took me about 80 hours to go from no knowledge to first patch. Assuming no knowledge, it is a two step process.
(1) Learn basic OS concepts through xv6.
xv6 is a reimplementation of an early version of Unix, designed to be as simple as possible and accompanied by a whole book of commentary. Get the book and the source cost listing printed and bound: http://pdos.csail.mit.edu/6.828/2012/xv6.html. Work through the book and exercises. Use the lecture videos from 6.828 from 2011 if you need extra material in order to understand: http://pdos.csail.mit.edu/6.828/2011/schedule.html.
(2) Pick a part of OSes that you are interested in. Contribute to that part in Linux.
Figure out where that part is in the Linux kernel. Find a bug in the bug tracker and submit the patch. I found filesystems interesting, so I fixed a small bug in one of the filesystems. Use a cross reference, it will save you a lot of time: http://lxr.free-electrons.com/source/include/linux/cpu.h. Also feel free to subscribe to the Linux kernel subreddit: http://www.reddit.com/r/kernel. I've set up the sidebar with a lot of useful links.
The Linux kernel is large and complex. You need to equipe yourself with a mental model of an OS through xv6 and then pick one small, specific part to attack in Linux. Be tactical! Otherwise you will be overwhelmed.
--
As an aside, I'm actually currently working on a tool that parses the Linux source code to find symbol definitions and then works its way back through the Git history to find the commit message for when the symbol was first defined. These commit messages usually contain really useful information about the original intent of the symbol and implementation details. Currently fighting with a few bugs in my C grammar, but should be able to work through those soon. Please feel free to email me at tsally@atomicpeace.com if you want to be pinged when the tool is released.
As an aside, I'm actually currently working on a tool that parses the Linux source code to find symbol definitions and then works its way back through the Git history to find the commit message for when the symbol was first defined.
Are you just doing something similar to cscope to find the definition of a symbol, then running git blame on that line? Or are you actually checking earlier revisions as well, to see if the symbol was moved or changed types?
I'm finding the definition of a symbol using a parsing expression grammar for C. I haven't yet decided upon the particular algorithm for finding the first commit, but I plan on using libgit2 to work directly with the repository. In the simplest case, you just iterate through all versions of a file and find the earliest one with the string in question. Obviously care is required when walking the repository history.
Thanks for the pointer. I know I'll find the commit using libgit2, but I hadn't yet gotten to thinking about the algorithm to do so. git log -S seems like a good starting point.
This is going to be unsatisfying, but the only thing that works for motivating me to understand a piece of kernel code has been wanting to change it and having to learn exactly how it works to achieve that.
If you just want to learn more about how the kernel fits together, reading http://lwn.net/Kernel/LDD3/ (Linux Device Drivers, freely downloadable) is a fine start.
This is the way the Linux kernel was introduced to us in Uni, and I really think it is the best way. Set a goal like "write a simple device driver", or (my choice) "implement a new concurrency primitive", and work towards that. Trying to do something like adding P/V system calls will teach you a lot about how the kernel operates; enough to give you a good starting place to find something new about it to learn.
I have this book and think it's great. It covers Linux 2.6, which happens to be the version of the kernel used in CentOS6.2 which is what I use at work. What are the major differences between 2.6 and the 3 line. Specifically, what is the major difference that caused the numbering scheme to go from 2 to 3?
One thing I found a little different was that the OS has its own libraries for everything (string.h etc..) which makes sense if you think about it.
If you want to browse the source code on-line with this software website thing called lxr (Linux Cross Reference). Its got a good search tool and linked headers. Clicking on a function name shows you where that function is used.
You can install it yourself and I think its much faster..
http://lxr.linux.no/linux+v3.6/http://rhkernel.org/
There is linux weekly news too, which is a decent site when I was still in the Linux porting world. Like many Linux sites, seems to lack in style, but makes up in content.
http://lwn.net/
start off with reading 1) linux kernel development 2) linux device drivers 3) linux kernel module programming guide
have - understanding the linux kernel as your reference manual.
by now, you should be comfortable to read/understand the kernel source; download linux kernel source and start browsing through the code.
simply reading books wont get you anywhere - you need to play around with kernel source inorder to understand the linux kernel behavior and different problems you may come across. write simple kernel modules to get a hang of how you can interact/modify with the kernel.
join some opensource project and start fixing bugs you're comfortable with or just play around with your local linux kernel source - make changes; build and deploy and observe what happens.
EDIT:
if you have no prior knowledge of OS Theory and Fundamentals; then you should start here first - read either of the following books
1) Operating System Concepts by Galvin, Silberschatz OR 2)Modern Operating Systems by Tanenbaum
For programming related - system calls and stuff
read 1) Advanced Programming in the UNIX Environment - by Richard Stevens
You may want to start with the earliest Linux kernels as they will be considerably smaller and simpler. You'll also be able to fit the whole kernel in your head.
From that point you can look at yearly diffs to find subsystems that have changed of interest. (Are 4 rewrites of USB interesting?)
Yeah I agree. The first versions of Linux were 10k LOC or less. Should be pretty readable.
One thing I just did was build the first version of git (just sync backwards in the git/git repo and type "make"). I ran the commands manually, and looked at the data file formats, and at the source code, and it greatly improved my understanding of git. It's all still relevant.
The first version of git was shockingly small, like 500 LOC of plain C code or so, but it does a surprising amount of the core work. I also gained a lot of respect for Linus' coding style.
People have said that Linux itself has too many hands in it -- e.g. the system call interface is a huge mess. So I wasn't sure if I would like Linus' code, but I definitely do after reading it.
I think he doesn't care about consistency when merging, because git's interface is a huge mess, just like the Linux syscall interface. But his code is consistent and good for sure.
There's no need to post shortened URLs to HN, and in fact this is discouraged. If you post a very long URL, HN will elide the text beyond some limit so it fits better in the page.
I'd really like to see btrfs start to stabilize and become more mainstream, but everything still seems to say that it's experimental.
The main reasons that I want it on my desktop are:
* easy backups
* checksums
* maybe it will become integrated with package manager (I'm on ubuntu) so that I can roll back package updates if they don't work properly
It is integrated with the package manager on Ubuntu already. I just upgraded my laptop to Quantal and part of the upgrade process spotted I was using btrfs and created a snapshot. (It was a complete waste of time as I have automatically created hourly, daily, weekly and monthly snapshots.) Doing regular package updates is very slow with btrfs. This is due to fsync() being called several times per file.
You'll also want Quantal for an updated btrfs-tools package. For example it lets you do scrubbing. You also need it (and kernel 3.3+) to change raid levels after filesystem creation. The Ubuntu 12.04 installer does create and install to btrfs but didn't let you configure things like raid level so you were stuck with whatever it did.
I've been running all my systems on btrfs for several months now on Ubuntu 12.04 (kernel 3.2) on both SSD and HDD including using RAID 0, RAID 1, dmcrypt/LUKS, inplace conversion from ext4 and who knows what else. (Across my server, workstation, laptop and HTPC.)
The only problem I have had is when filesystems fill up. I've never lost data but it can be quite frustrating trying to find files to delete (also need to be removed from snapshots), rebalancing etc to get things running smoothly again. The various ways of doing df are mostly a work of fiction.
My data is duplicated to offline backups, Dropbox, multiple systems and git/hg repositories so the failure of any system would be annoying but I'd never lose anything. You should make sure you are in that position first, independent of what filesystems are being used.
The former - fsyncs are more expensive because they involve copy on write updates all the way up the tree to the root. Some of the fsyncs are there because of ext4 quirks. That all said, every kernel release brings big improvements in btrfs especially metadata handling. IOW it is currently the worst it will be.
Why not just sync the data change and sync the metadata changes only to a write-ahead log? Or if the data is logged also, you only need to sync the log.
You can install apt-btrfs to get snapshots when installing thingsm and run "eatmydata apt-get install <x>" to remove fsyncing. Of course, you would be out of luck if the power went out in the middle of the installation, but it makes things much faster.
You can revert to the snapshot so there isn't really a problem, although that can be quite finicky and probably take longer to setup than the time saved avoiding fsyncs.
Note that btrfs will be just fine on unexpected power outage - the filesystem will just contain some random combination of old and new files affected by the installation.
The fsck for btrfs is still not quite there. If you are using multiple disks with btrfs's mirroring/checksums, hopefully you won't get hit by HDD hiccups leading to metadata corruptions which btrfs's fsck can't yet handle. I certainly wouldn't recommend running btrfs on a single HDD/SSD unless you are doing frequent backups and don't mind losing state between backups if you get unlucky.
Also, please note that btrfs snapshots != backups. They will not save you in case of device failure. Checksums also won't help you if there are kernel bugs.
I have been running Ubuntu 12.04 with my / as btrfs for a couple of months now, with no trouble (and not even thinking about it truth be told). That said, I left my /home ext4 because I didn't know how much to trust it.
Did the same, btrfs for everything but home and will switch to ext4 with 12.10 because it feels much slower then the ext3 i had earlier.
But i'll be sure when i upgrade.
To be honest, what do i need btrfs for? I don't (and likely wouldn't) use all of its advanced features anyway..
In all the benchmarks ext4 seems to be a reasonable choice in terms of speed and stability.
I am looking forward to when we have fire and forget snapshotting system so that we all have rolling backups. I would also like the send snapshot features they have added recently for keeping multiple drives in sync (I currently use unison, but I think this could be faster/better).
Plus copy on write should be pretty cool (cp becomes as fast as mv), though I could see it allowing me to engage in some suboptimal behaviors of copying tons of stuff and counting on CoW to avoid actual duplication.
And I still can't sleep my Powerbook /w Radeon because post KMS radeon doesn't support it yet (no one ported the pre-KMS code from the X driver portion of radeon into the post-KMS radeon kernel driver).
OTOH, I can just leave KMS off and have no 2D or 3D acceleration.
The code is already out there, all 3.7 integration would represent is a stamp of approval. Consider that there is no hardware for the architecture yet, I don't know why approval would be very interesting:
From the desktop perspective, I really like the "Suspend to disk and memory at the same time" part. It's great that you don't have to think anymore if you're going to use your laptop in the next 2 days or not. If it runs out of power, it does - nothing gets lost!
Adding bufferbloat pieces into the main release is also good news (although that probably won't hit the servers that really need it for many more months...)
Do you think this will obsolete the 'hybrid suspend' functionality in pm-utils? I think that works by suspending to ram, waking up in 10 minutes then immediately suspending to disk.
Maybe a combination of these two approaches could be useful.
I mean, I guess I can just dive in, and maybe that's the best approach, but is there a strategy anyone would recommend for reading the Linux source in terms of it making sense as a combined unit of code (as opposed to a collection of algorithms, if that makes sense)?