Re: Latest brk patchset

[Jeremy Fitzhardinge <jeremy@goop.org>]

H. Peter Anvin wrote:
> Jeremy Fitzhardinge wrote:
>   
>>> It really doesn't make much sense to me, and is more than a bit
>>> confusing given the symbols.
>>>       
>> Mostly because I knew that the bss would get mapped into the appropriate
>> phdr segment correctly, but I wasn't sure that another bss-like section
>> would be.
>>     
>
> It will; in fact if they are adjacent then ld will typically merge the
> PHDRs.
>
>   
>> Also because historically the brk segment was just an
>> extension of the executable's bss, and its more or less the same too.
>>     
>
> An extension of, yes, but not a part of.
>
>   
>> Is there any real benefit in putting it into another section?
>>     
>
> Well, the semantics are different; the .bss section is zeroed while the
> brk isn't,
Traditionally, brk is always zeroed.  extend_brk() zeros the memory it 
returns (to be consistent with bootmem, and to make it easier to migrate 
from bss -> brk).

>  and the brk symbols don't necessarily point to the data
> associated with those particular symbols, unlike (of course) the bss.
>   

Yes, its a bit of a pitfall.  I guess the symbols are useful as a way to 
identify brk users just from looking at the vmlinux, but they're not 
really all that useful.  I'm half thinking we should put some non-C 
identifier characters in them to make sure that C code can never refer 
to them.

    J