Re: [PATCH 1/4] jbd2/journal_commit_transaction: relocate state lock to incorporate all users

[Theodore Ts'o <tytso@mit.edu>]

On Mon, Jun 10, 2013 at 10:45:50PM -0400, Paul Gortmaker wrote:
> 
> Sure, I will do so tomorrow -  but since it can't be reproduced
> on-demand, all I'll be able to do is to watch for independent
> calls with very close time stamps, and confirm they were not
> interleaved.

Well, if the resulting patch causes jbd_debug() messages to be issued
correctly, I have every confidence that they won't be interleaved; the
%pV structure was used to solve this very problem, and it's used all
over the kernel.  In fact we've used it __ext4_error()/ext4_error() to
solve this exact same issue of interleaved messages.

> What about the state assert being done outside of the state
> lock?   Should I keep that as a separate patch so that the
> assert isn't checking what could possibly be a transient value?

Ah, I missed that since I had been focusing on the jbd_debug().
That's a good catch, we'll still need this patch to make sure we're
checking the state assert under the j_state lock.  But with the fixed
jbd_debug() we can keep the jbd_debug() statement outside of the
j_state_lock critical region.

Cheers,

						- Ted