Re: Definitions

[Andi Kleen]

On Thu, Aug 10, 2000 at 02:21:09PM +0200, Jamie Lokier wrote:
> Andi Kleen wrote:
> > In practice the memory pinning does not seem to be a problem with reiserfs.
> > Reiserfs currently always uses a 32MB log which limits the pinned memory
> > to 32MB per file system (all pinned buffers need to have log space
> > reserved).
> 
> Ouch.  I need 32MB per file system?  That's more RAM than my 128MB
> machine has...  and 32MB is way more RAM than the last 486 I used.

No, you don't need it. It is just what could be pinned in the worst case
in case you manage to generate that much dirty metadata quickly enough
that the flushing cannot cope.

Note that TCP (socket buffers) and other subsystems like fork() (process
stacks) or the mm subsystem (page tables) can use much more pinned memory, 
so better stop using all these evil things too. 

> 
> > In addition it only logs metadata and you have to do a lot of work to
> > quickly dirty 32MB or more of metadata. It is rather unlikely in
> > practice because you would need to deal with thousands of files to
> > create that much dirty metadata.
> 
> find /  ->  100,000 dirty atime inodes.

Which are clustered in much less blocks and flushed lazily in small parts
(they are only pinned while they are flushed) 

-Andi