Re: PROBLEM: ramfs reports 0 free

[Willy Tarreau <willy@w.ods.org>]

On Sat, Apr 29, 2006 at 07:52:16PM +0000, Senior Goat wrote:
> [1.] One line summary of the problem:
> 
>    ramfs erroneously reports 0 bytes free, which confuses some programs.
> 
> [2.] Full description of the problem/report:
> 
>    The system call on a mounted ramfs, as indicated by df, reports 0 
>    bytes total, used, and free.  If ramfs is expected to perform like a 
> filesystem, it should not do this.

(...)

> [X.] Other notes, patches, fixes, workarounds:
> 
>   I read that Linus himself wrote this module.  For some reason he 
>   decided to report 0, but I can't figure why.  Perhaps the overhead for 
> finding information was too great.
> 
>  Couldn't you add up the amount of filesystem cache with the free memory 
> and get a crude, but quick estimate of the amount of free space available 
> for any given ramfs.
> 
>  I'm not sure how to handle the total space, since you probably don't 
>  want that fluctuating too much, except that you might just report the 
> total amount of ram(which won't fluctuate), and then report the used ram. 
> This is all information that /free/ reports from system calls with little 
> delay.
>
>  The only other place I can think that ramfs might get memory is when the 
> kernel swaps out other processes, but you can't count on that.
> 
>  To sum it up, the best way to get a semi-valid report would be:
> total:  total ram
> used: used ram (which takes into account memory used in the ramfs, 
> coincidentally)
> free:  total - used  (which ignores disk cache, since that should be 
> freed when needed)

To achieve this, you have to set an arbitrary limit on the maximal FS size.
I have updated a patch originally from David Gibson to enforce limits on
RAMFS. His original patch for 2.4.19-rc1-ac2 as well as my update for
2.4.32 is available here, in case you're interested in porting it to 2.6 :

   http://w.ods.org/linux/kernel/2.4/lkup/ramfs-limit.html
   
>   The other option would be to make all the little programs (like 
>   Debian's package manager) check if the filesystem it wants to write to is 
> a ramfs before reporting an error, but this is a blatant hack.

This is very hard to maintain, that's why I include the patch above in
my kernels ;-)

Cheers,
Willy