lkml.org 
[lkml]   [2005]   [Oct]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] fix nr_unused accounting, and avoid recursing in iput with I_WILL_FREE set
On Tue, Oct 18, 2005 at 06:15:48PM -0700, Andrew Morton wrote:
> > > Well according to my assertion (below), the inode in __sync_single_inode()
> > > cannot have a zero refcount, so the whole if() statement is never executed.
> >
> > generic_forget_inode->write_inode_now->__writeback_single_inode->
> > __sync_single_inode
>
> oshit.

When does this ever happen? Just for private inodes released during
put_super right?

>
> > We do have I_WILL_FREE, but i_count will be zero.
>
> yup.
>
> > >
> > > The thinking behind that increment is that __sync_single_inode() has just
> > > taken a dirty, zero-refcount inode and has cleaned it. A dirty inode
> > > cannot have previously been on inode_unused, hence we now are newly moving
> > > it to inode_unused.
> >
> > nr_unused doesn't seem to count the number of inodes on the unused list.
> > It is actually counting the number of inodes whose i_count is 0. See
> > generic_forget_inode and invalidate_list to see what I mean.
>
> 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.

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? 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.

>
> --- devel/fs/inode.c~generic_forget_inode-nr_unused-cleanup 2005-10-18 18:13:22.000000000 -0700
> +++ devel-akpm/fs/inode.c 2005-10-18 18:13:57.000000000 -0700
> @@ -1067,8 +1067,8 @@ static void generic_forget_inode(struct
> if (!hlist_unhashed(&inode->i_hash)) {
> if (!(inode->i_state & (I_DIRTY|I_LOCK)))
> list_move(&inode->i_list, &inode_unused);
> - inodes_stat.nr_unused++;
> if (!sb || (sb->s_flags & MS_ACTIVE)) {
> + inodes_stat.nr_unused++; /* One more 0-ref inode */
> spin_unlock(&inode_lock);
> return;
> }
> @@ -1077,7 +1077,6 @@ static void generic_forget_inode(struct
> write_inode_now(inode, 1);
> spin_lock(&inode_lock);
> inode->i_state &= ~I_WILL_FREE;
> - inodes_stat.nr_unused--;
> hlist_del_init(&inode->i_hash);
> }
> list_del_init(&inode->i_list);
> _
>
> > generic_forget_inode took care of incrementing the unused count when
> > i_count went to zero. So, I don't think we need to worry about the
> > unused count in __writeback_single_inode.
> >
>
> How about this for now?

This part looks good.

-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/

\
 
 \ /
  Last update: 2005-10-19 04:01    [W:0.982 / U:0.028 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site