Up until the early 2000s I used to compile my own kernel, carefully selecting only the options that I needed.
Then I realised that I wasn’t saving memory, because almost everything was a module anyway.
Is there any actual benefit to using a custom kernel on consumer hardware that’s supported by the stock kernels?
Your question is one of the two reasons I love GNU/Linux so much, and will not go back to proprietary tools ever again :)
Even tough as a very average user myself I would never feel like compiling my own kernel, and would even less know how to do such a thing, I know it’s a possibility and I know other users are doing it. And that is a possibility only because of the freedoms we the users are given by the GPL to do… what we want. To me, as an ex-lifelong Apple user (I started being their customer in the early 80s and only switched full time some 7 or 8 years ago to GNU/Linux) this is amazing and wonderful freedom.
Sorry if I have not replied precisely to your question but reading it I realized it was a great demonstration of what freedom is supposed to mean, and I felt like sharing it.
You could compile with
-march=nativeto get an optimised build, but its unlikely to show any benefit outside benchmarks.It’s a matter of pride, honor, and the classical ideals of antiquity!
For consumer hardware supported by stock kernels? No advantage at all. At most you may want to switch kernels, but most distros have a handy tool for that.
The only time I’ve compiled my own kernel in the last 15 years has been for work on very specialized embedded systems.
Actual benefit? I’d say code not compiled in is code that cannot be abused.
There are also more obscure benefits - like getting more familiar with your system.
Primarily if you want some functionality that isn’t mainlined, or isn’t released as stable yet.
Like hibernate in lockdown mode, or out of tree drivers, or maybe something new coming up in the emulation support world like NTSync, though I think that last example was mainlined by now.
I think that last example was mainlined by now.
Yeah NTSync is in kernel 7.0x
Woah, I didn’t know they were working on those features. Thanks for sharing!
Drivers are usually loaded as modules, aren’t they?
I’m not that certain to be honest. But the following is my best understanding:
-
Most drivers are included directly in the kernel source.
-
They can be compiled built-in directly to the kernel binary, or they can be compiled as loadable kernel modules. I don’t know how the proportions between the two options look, but at least the essential drivers (chipsets, filesystems, etc) should be compiled in to allow the boot to progress enough that module loading works.
-
There are some, like the Nvidia proprietary GPU driver, that are provided only in binary form as loadable kernel modules.
-
I also understand that a lot of smartphone drivers are developed out-of-tree against older branches of the Linux kernel. Even those that are made public / open sourced, end up living outside the mainline kernel, and the devs of third party android builds have to cherry pick them into their kernel source.
I think at least the last group should count as an example of a reason of the type for what OP was looking for.
-
I’ve only done it to REMOVE features but that’s just a silly exercise in making the absolute tiniest OS I can
Normal user? Extremely rarely would you need to build the kernel. Distributions design their options to fit most use cases, and you’ve observed the extensibility through modules. The kernel itself has moved towards runtime configurable options for your convenience over time, such as with
PREEMPT_DYNAMICWhere in the past changing the preemption model would require a recompile. Ultimately, this is a good thing; it makes your life easier and you can get better support for a common kernel if you need to debug.
It does happen though if you need special hardware or if you’re picky about specific kernel features. For example, I’ve used kernels that don’t have built-in support for memory compression. Need is a subjective term, and I felt that was a configuration option that I needed because a memory upgrade was not an option. I would argue there was a point to that effort. Considering that you phrase your question as asking about normal users, then no, I would say that’s rarely beneficial, might actually be disadvantageous because you won’t receive as much help debugging problems from your distribution, and generally you can achieve your goals by tuning runtime kernel parameters anyway.
In all my time with Linux, I have never once recompiled. I’d go as far to say as it is never necessary for a normo to do it.
The only time I have compiled a kernel in the last 10 years was to bisect to determine what change introduced a bug affecting my particular hardware combination.
Not really unless you need a specific optimization or module that isn’t available otherwise. Most distros make distribution of external modules available via package manager, and most of the optimizations you would want to enabke can be turns on or off elsewhere as feature flags.
