Re: jbd2_handle and i_data_sem circular locking dependency detected

["Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>]

On Mon, Feb 04, 2008 at 05:31:56PM +0100, Jan Kara wrote:
>   Hi,
> 
> On Mon 04-02-08 15:42:28, Aneesh Kumar K.V wrote:
> > This is with the new ext3 -> ext4 migrate code added. The recently added
> > lockdep for jbd2 helped to find this out. We want to hold the i_data_sem
> > on the ext3 inode during migration to prevent walking the ext3 inode
> > when it is being converted to ext4 format. Also we want to avoid 
> > file truncation and new blocks being added while converting to ext4.
> > Also we dont want to reserve large number of credits for journal.
> > Any idea how to fix this ?
>   Hmm, while briefly looking at the code - why do you introduce i_data_sem
> and not use i_alloc_sem which is already in VFS inode? That is aimed
> exactly at the serialization of truncates, writes and similar users.
> That doesn't solve problems with lock ordering but I was just wondering...
>   Another problem - ext4_fallocate() has the same lock ordering problem as
> the migration code and maybe there are others...
>   One (stupid) solution to your problem is to make i_data_sem be
> always locked before the transaction is started. It could possibly have
> negative performance impact because you'd have to hold the semaphore for
> a longer time and thus a writer would block readers for longer time. So one
> would have to measure how big difference that would make.
>   Another possibility is to start a single transaction for migration and
> extend it as long as you can (as truncate does it). And when you can't
> extend any more, you drop the i_data_sem and start a new transaction and
> acquire the semaphore again. This has the disadvantage that after dropping
> the semaphore you have to resync your original inode with the temporary
> one your are building which probably ends up being ugly as night... Hmm,
> but maybe we could get rid of this - hold i_mutex to protect against all
> writes (that ranks outside of transaction start so you can hold it for the
> whole migration time - maybe you even hold it if you are called from the
> write path...). After dropping i_data_sem you let some readers proceed
> but writers still wait on i_mutex so the file shouldn't change under you
> (but I suggest adding some BUG_ONs to verify that the file really doesn't
> change :).
> 

This one came up again when i was doing the mmap ENOSPC patch. Now the
current code with migrate is taking i_mutex to protech against all
writes. But it seems a write to a mmap area mapping a hole can still go
through fine. And that path cannot take i_mutex.

So i guess the migrate locking need to be fixed. Any suggestion ?

-aneesh