Re: next-20090310: ext4 hangs

From: Jiri Gaisler <jiri@gaisler.com>
To: Jan Kara <jack@suse.cz>
Cc: Alexander Beregalov <a.beregalov@gmail.com>,
	Theodore Tso <tytso@mit.edu>,
	"linux-next@vger.kernel.org" <linux-next@vger.kernel.org>,
	linux-ext4@vger.kernel.org, LKML <linux-kernel@vger.kernel.org>,
	sparclinux@vger.kernel.org
Subject: Re: next-20090310: ext4 hangs
Date: Thu, 26 Mar 2009 01:17:06 +0100	[thread overview]
Message-ID: <49CAC982.4050507@gaisler.com> (raw)
In-Reply-To: <20090326000035.GB11024@duck.suse.cz>

Jan Kara wrote:
>   Hi,
> 
>   thanks for the updated disassembly.
> 
>>   55bab0:       82 10 60 01     or  %g1, 1, %g1
>>                 spin_unlock(&journal->j_list_lock);
>>   55bab4:       40 06 4b 20     call  6ee734 <_spin_unlock>
>>   55bab8:       c2 24 e0 28     st  %g1, [ %l3 + 0x28 ]
>   OK, so it really seems that:
>   jinode->i_flags |= JI_COMMIT_RUNNING;
>   spin_unlock(&journal->j_list_lock);
> 
>   has been compiled to
>   ld [ %l3 + 0x28 ], %g1
>   or  %g1, 1, %g1
>   call  6ee734 <_spin_unlock>
>   st  %g1, [ %l3 + 0x28 ]
> 
>   Which seems like a bug in the compiler or in the way implement compiler
> barriers in spin_unlock() on UP sparc. Or is there some sparc magic by which
> this is correct code? Any clever sparc guy?
> 
> 									Honza

Looks fine to me - the store is made in the branch delay slot
of the call, so it will occur before the first instruction at
the call target address executes.
Jiri.