Re: DIO process stuck apparently due to dioread_nolock (3.0)

[Michael Tokarev <mjt@tls.msk.ru>]

12.08.2011 21:01, Eric Sandeen wrote:
[]
>> And yes, 18446744073709551199 aios sounds quite alot ;)
> 
> looks like it went negative.
> 
> I see that in one case we set EXT4_IO_END_UNWRITTEN, but don't increment the counter.
> We decrement the counter for every EXT4_IO_END_UNWRITTEN completion, I think.
> 
> I'm not quite sure if that was intentional or not, but it might be a place to start.
> I haven't though hard about this, in the middle of something else right now,
> but this looks like it's a probllem in my code from that unaligned AIO patch,
> perhaps...
> 
> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> index 3e5191f..7366488 100644
> --- a/fs/ext4/inode.c
> +++ b/fs/ext4/inode.c
> @@ -3640,6 +3640,7 @@ static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
>  
>         io_end->flag = EXT4_IO_END_UNWRITTEN;
>         inode = io_end->inode;
> +       atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
>  
>         /* Add the io_end to per-inode completed io list*/
>         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);

This patch does not change thing (unfortunately I forgot to reapply
the debugging patch posted by Jan so don't know how many aiocbs it
wants to wait).

But reverting e9e3bcecf44c04b9e6b505fd8e2eb9cea58fb94d
(ext4: serialize unaligned asynchronous DIO) helps, the thing
goes fine without that patch.

And this refcounting - I can't see why in my case the problem only
happens with hot cache and only with dioread_nolock mount option
(so far anyway - I weren't able to trigger it without at least
one of the two conditions).

BTW, ext4_end_io_buffer_write() is declared twice in fs/ext4/inode.c

Thanks,

/mjt