Re: [PATCH] fix nr_unused accounting, and avoid recursing in iput with I_WILL_FREE set

[Chris Mason <mason@suse.com>]

On Tue, Oct 18, 2005 at 07:26:46PM -0700, Andrew Morton wrote:
> > > hm, OK.  It'd be nice to make that more explicit.  Something like this?
> > 
> > Well, I can't quite convince myself it is wrong, but when 
> > (!sb || (sb->s_flags & MS_ACTIVE), we're dropping the
> > inode_lock with an inode with i_count == 0 and nr_unused hasn't been
> > incremented.
> > 
> > So, if someone (sync_sb_inodes?) comes in and runs __iget,
> > the counts end up wrong.  Then again, whoever ran __iget would also run
> > iput and things would go horribly wrong anyway.
> 
> Nope, it's equivalent:

The math ends up the same, but for your version there is a window where
the lock isn't held and the count doesn't reflect reality.  I don't know
if anyone can race in and mess with the inode though.  It is still on
various lists, but if we're only in that part of generic_forget_inode
during unmount, the super semaphore will keep out most potential racers.

I need to read harder.

> > Did I mention the part where Andrea and I are hunting a bug where the
> > count of unused inodes goes negative and the everyone ends up spinning
> > in shrink_icache_memory?
> 
> No.
> 
> >  Andrea's patch doesn't fix the spinning, but
> > it might have fixed the unused inode count going negative.  We're
> > waiting for another reproduce on the ppc64 race monster.
> 
> I assume you have BUG_ON(inode_stat.nr_unused < 0)s in there everywhere?
> 
> In fact WARN_ON(inode_stat.nr_unused < 100) might be better - something's
> obviously doing a bogus decrement a lot of times.
> 

It goes negative in the invalidate_inodes run during unmount.  I
think Andrea's patch will solve that part, hopefully we'll know more
tomorrow.

-chris

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/