Re: inode_unused list corruption in 2.4.26 - spin_lock problem?

[Marcelo Tosatti]

On Mon, Jun 21, 2004 at 01:10:21PM -0400, Trond Myklebust wrote:
> På su , 20/06/2004 klokka 20:45, skreiv Marcelo Tosatti:
> > Lets see if I get this right, while we drop the lock in iput to call 
> > write_inode_now() an iget happens, possibly from write_inode_now itself 
> > (sync_one->__iget) causing the inode->i_list to be added to to inode_in_use. 
> > But then the call returns, locks inode_lock, decreases inodes_stat.nr_unused--
> > and deletes the inode from the inode_in_use and adds to inode_unused. 
> > 
> > AFAICS its an inode with i_count==1 in the unused list, which does not
> > mean "list corruption", right? Am I missing something here?
> 
> Yes. Please don't forget that the inode is still hashed and is not yet
> marked as FREEING: find_inode() can grab it on behalf of some other
> process as soon as we drop that spinlock inside iput(). Then we have the
> calls to clear_inode() + destroy_inode() just a few lines further down.
> ;-)
> 
> If the above scenario ever does occur, it will cause random Oopses for
> third party processes. Since we do not see this too often, my guess is
> that the write_inode_now() path must be very rarely (or never?) called.

Thats what I though: That if the scenario you described really happens, we 
would see random oopses (processes using a deleted inode) instead of 
Chris's list corruption.

Chris, _please_ post your full oopses.

> > If you are indeed right all 2.4.x versions contain this bug.
> 
> ...and all 2.6.x versions...
> 
> I'm not saying this is the same problem that Chris is seeing, but I am
> failing to see how iput() is safe as it stands right now. Please
> enlighten me if I'm missing something.

For me your analysis looks right and we have a problem here.

I think Al Viro knows iput() very well. Maybe he should take a look 
at your patched. CC'ed.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/