Re: VM deadlock

[Chris Mason]

On Wednesday, June 27, 2001 04:43:28 PM -0300 Rik van Riel
<[e-mail]> wrote:

> On Wed, 27 Jun 2001, Chris Mason wrote:
> 
>> 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.
> 
> If you don't have free memory, you are limited to 2 choices:
> 
> 1) wait on IO
> 2) spin endlessly, wasting CPU until the IO is done
> 
> If (1) isn't possible in reiserfs, I'd say something in
> reiserfs needs to be fixed, otherwise you will always
> have problems when the system has lots of dirty mappings
> that need to be written out.
> 

Ok, I need to describe the problem a little better.  reiserfs inodes need
to be logged, which means you have to join/start a transaction in order to
write them.

So, if kswapd tries to write them, it might end up waiting on the log.
Normally this is not a big deal, but almost allocations in reiserfs use
GFP_BUFFER, which means we never end up doing i/o ourselves in
page_launder, and always end up waiting on kswapd.  So, kswapd waits on
reiserfs and reiserfs waits on kswapd (none of these are spin locks ;-)

The work around I've been using is the dirty_inode method.  Whenever
mark_inode_dirty is called, reiserfs logs the dirty inode.  This means
inode changes are _always_ reflected in the buffer cache right away, and
the inode itself is never actually dirty.

So, the only time reiserfs_write_inode needs to do something is for fsync
and/or O_SYNC writes, and all it needs to do is commit the transaction.  

Any time kswapd is calling write_inode, it is just trying to free the inode
struct, and reiserfs can safely ignore the write request, regardless of if
a sync is requested.

-chris