[dm-devel] Re: [PATCH] drop mutex in __set_size

[Mikulas Patocka]

On Wed, 29 Jul 2009, Jun'ichi Nomura wrote:

> Hi Mikulas,
> 
> Mikulas Patocka wrote:
> > Drop the mutex.
> > 
> > It doesn't make sense to lock it for a single assignment, this code can't
> > be executed concurrently. The size should be read with i_size_read which
> > is automatically protected against concurrent i_size_write.
> 
> But it seems there are codes which depend on i_mutex for
> protected access to i_size: e.g. block_write_begin().

You are right, but I don't know what to do with it. Catch such uses and 
convert them to i_size_read?

> And while my description only mentioned noflush suspend,
> with 2.6.31-rc, the deadlock can occur under normal suspend
> because the new barrier code may push bios to deferred queue.

It can deadlock with any code that takes i_mutex and submits or waits for 
i/os, regardless of barriers.

Mikulas