Re: [PATCH V5 2/2] btrfs: implement delayed inode items operation

[Chris Mason <chris.mason@oracle.com>]

Excerpts from Miao Xie's message of 2011-03-27 07:44:06 -0400:
> On sun, 27 Mar 2011 20:09:10 +0900, Itaru Kitayama wrote:
> > Hi Miao,
> > 
> > On Sun, 27 Mar 2011 15:00:00 +0800
> > Miao Xie <miaox@cn.fujitsu.com> wrote:
> > 
> >> I got it. It is because the allocation flag of the metadata's page cache, which is stored in
> >> the btree inode's i_mapping, was set to be GFP_HIGHUSER_MOVABLE. So if we allocate pages for
> >> btree's page cache, this lockdep warning will be triggered.
> >>
> >> I think even without my patch, this lockdep warning can also be triggered, btrfs_evict_inode()
> >> do the similar operations like what I do in the btrfs_destroy_inode(). 
> >>   Task1                    Kswap0 task
> >>   open()
> >>     ...
> >>     btrfs_search_slot()
> >>       ...
> >>       btrfs_cow_block()
> >>     ...
> >>     alloc_page()
> >>       wait for reclaiming
> >>                     shrink_slab()
> >>                       ...
> >>                       shrink_icache_memory()
> >>                         ...
> >>                         btrfs_evict_inode()
> >>                           ...
> >>                           btrfs_search_slot()
> >>
> >> If the path is locked by task1, the deadlock happens.
> > 
> > Ok. balance_pgdat() calls shrink_slab() with GFP_KERNEL so it's still possible for the kswapd0 
> > to call prune_icache(), no? I still see the lockdep warning even with your patch that clears
> > __GFP_FS in open_ctree().
> 
> sorry for my mistake. The above explanation is wrong, it has no business with kswap thread.
> The correct explanation is
> 
>    Task1
>    open()
>      ...
>      btrfs_search_slot()
>        ...
>        btrfs_cow_block()
>      ...
>      alloc_page()
>        do_try_to_free_pages()
>          shrink_slab()
>            ...
>            shrink_icache_memory()
>          ...
>          btrfs_evict_inode()
>            ...
>            btrfs_search_slot()
> 
> If the path is locked by task1, the deadlock happens.
> 
> So balance_pgdat() is impossible to trigger the lockdep.
> (My clearing __GFP_FS patch's changelog is also wrong.)
> 
> I see, except btree's page cache, free space cache's page cache is also special,
> can not use __GFP_FS flag.

Ok, I've got your first patch already, I'll add a hunk for the free
space cache too.  Most of the allocations we're doing are explicitly
with GFP_NOFS, so it is just supporting allocations and readahead that
should be causing trouble.

Thanks!

-chris