Skip to contentSkip to conversation

Less generic kernels

DISCUSSION Ralf Baechle 1 participant 1 message Dec 4, 2003 MIPS

lore
Share & tools
lore
#0 lore raw

A side effect of cleaning the processor specific code a while ago was that every system was carrying every bit of cpu-specific code such as cache code with it that Linux only had to offer. I cleaned that, this are the savings for two default configurations:

text data bss dec hex filename

2388266  275076   92456 2755798  2a0cd6 vmlinux         defconfig	old
2385522  275076   92456 2753054  2a021e vmlinux         defconfig	new

-> 2744 bytes saved

2197189  641168  128048 2966405  2d4385 vmlinux         defconfig-ip27	old
2190310  641168  128048 2959526  2d28a6 vmlinux         defconfig-ip27	new

-> 6879 bytes saved

Of course this also means faster, more efficient code.

How to make use of this for a particular system?

Just provide a include/asm-mips/mach-<system>/cpu-feature-overrides.h file. Any values you don't define in that file will receive default values from include/asm-mips/cpu-features.h.

That means if a particular macro isn't a constant for a particular system, just don't define the cpu_* macro that tests for it. Better even, if you don't know every little detail about your processor, just don't define the test macro and Linux will

Who gains most?

Obviously the most restrictive configurations gain most. Generally that's the case for typical embedded systems. Most of the size reduction can be achieved by knowing the cache line size in advance. This means the kernel will only carry one version for a particular processor around.

Oh, and this is a 2.6 only ...

Ralf