Re: VM deadlock

[Chris Mason]

On Wednesday, June 27, 2001 02:43:57 PM -0300 Marcelo Tosatti
<[e-mail]> wrote:
> 
> Looking at http://lists.omnipotent.net/reiserfs/200106/msg00214.html:

Also from Xuan ;-)

> 
>>> EIP; c0128228 <page_launder+b8/90c>   <=====
> Trace; c01303df <refill_freelist+1f/54>
> Trace; c01307e2 <getblk+f2/108>
> Trace; c5141308 <END_OF_CODE+4e978b8/????>
> Trace; c0176c4b <do_journal_end+63f/ac0>
> Trace; c5160848 <END_OF_CODE+4eb6df8/????>
> Trace; c01759e6 <journal_end_sync+16/1c>
> Trace; c015e23a <reiserfs_write_inode+56/64>
> Trace; c0141055 <try_to_sync_unused_inodes+101/1a8>
> Trace; c01416dd <prune_icache+105/114>
> Trace; c014170d <shrink_icache_memory+21/30>
> Trace; c0128d67 <do_try_to_free_pages+2b/58>
> Trace; c0128deb <kswapd+57/e4>
> Trace; c0105434 <kernel_thread+28/38>
> 
> 
> 
> refill_freelist() calls page_launder(GFP_BUFFER). Now GFP_BUFFER _will_
> block writting out buffers with try_to_free_buffers().

Grrr, how did I miss this before?  I thought Xuan's hang went away after
pre3, so I didn't look into this trace hard enough.  

Reiserfs expects write_inode() calls initiated by kswapd to always have
sync==0.  Otherwise, kswapd ends up waiting on the log, which isn't what we
want at all.

The dirty inode callback ensures there are no dirty inodes that haven't
been logged.  I took the sync parameter to mean it is initiated by fsync or
O_SYNC, so I trigger a full commit when sync == 1.

So, my choices are to ignore sync == 1 write_inode calls when kswapd is
doing it, or make a private inode dirty list.

> 
> Maybe thats the reason for the deadlock we're seeing here at this specific
> trace ? 
> 

The trace above is caused by the dirty inode problem, the I think the more
recent trace is something different.

-chris