__initstr & __exitstr

Hi,

Please comment on this approach to move strings in __init functions
from .rodata to .data.init so that it get discarded after initialization, like the variables marked as __initdata and the functions marked as __init, as well as move strings in __exit marked functions to .data.exit, that will be discarded and not even get into the generated kernel image.

Please note that if possible the best approach was for gcc to move
those strings automatically if the function was marked with modified __init/__exit macros, but we have to keep in mind that some of the strings in those functions can not be discarded because they keep being referenced by say register_chrdev and others, unlike, for example, proc functions and others that copy the string passed to some malloc'ed data structure, so we have to be selective marking exactly the ones that can indeed be discarded.

I've also implemented helper functions for printk thats the most
common case, and leaved the other common case, panic, using __initstr/__exitstr explicitely, so that people can comment on what is better.

Here is the basic implementation in include/linux/init.h:

#define __initstr(s) ({ static char __tmp_init_str[] __initdata=s;
__tmp_init_str;})
#define __exitstr(s) ({ static char __tmp_exit_str[] __exitdata=s;
__tmp_exit_str;})
#define init_printk(fmt,arg...) printk(__initstr(fmt) , ##arg) #define exit_printk(fmt,arg...) printk(__exitstr(fmt) , ##arg)

For modules its a no op, as modules doesn't get rid of code/data
marked as __init{data}, please correct me if I'm wrong as I didn't checked that in detail, but from first quick analysis it doesn't move it to some different .data/.text section, so I assume it doesn't discards it after initialization.

For my config, compiling everything statically, with a pristine
2.4.6-ac1 kernel I get 172 KB freed after init, with this patch we save 16 KB more.

I've put the patch at:
http://bazar.conectiva.com.br/~acme/patches/wip/__initstr.patch.4

And yes, its intrusive, but it serves, IMHO, as an experiment to
see how much can be saved with this.

Please advise/comment.

- Arnaldo

.. or fix all registration functions to use a private copy of the string, which would avoid some common oopses.

I dislike init_printk; it combines variadic functions/macros with assumptions about how the first argument is specified (i.e. as a string constant), so it's potentially very confusing.

Also, printk is used for debugging, and accidentally using init_printk instead of printk would result in no messages being printed at all if the driver is compiled into the kernel (while everything would work fine if the driver is compiled as a module, where init_printk and printk are identical). I think this would be very annoying to track down for less experienced kernel hackers.

I think this would fail if used in structure initialisations ?

Philipp Rumpf

Em Fri, Jul 06, 2001 at 10:05:14PM -0500, Philipp Rumpf escreveu:

yes, that would be nice, if allowed I can put my janitor hat and do that :)

that's why I've didn't coded a init_panic :) ok, I can change this if the rest of the patch is considered ok

yes, this fails in such cases, any ideas on another approach that works in this case as well in the other cases covered by the patch?

Also please note that strings in __initdata marked structs with just 'char *' instances are going to .rodata and not getting discarded, thats why I've changed some to char[SOME_MAX_STRING_SIZE], so that it gets moved to .data.init, crude, but works, I'm interested in another approach for this case as well.

Thanks for the comments.

- Arnaldo