Re: [PATCH -stable] sysfs: disable reclamation by default

[Maneesh Soni <maneesh@in.ibm.com>]

On Thu, May 17, 2007 at 07:49:31PM +0200, Tejun Heo wrote:
> Maneesh Soni wrote:
> > On Thu, May 17, 2007 at 05:04:23AM -0700, Greg KH wrote:
> >> On Wed, May 16, 2007 at 08:31:00PM +0200, Tejun Heo wrote:
> >>> sd->s_dentry updates made by dentry/inode reclamation are racy and can
> >>> lead to BUG() or oops.  This is already fixed in -mm and the fix is
> >>> scheduled to be merged into upstream for 2.6.23 but the fix
> >>> reimplements sysfs dentry dropping and is too risky for -stable
> >>> kernels.
> >>>
> > 
> > But was the synchronization fix tested by people facing the race? I still
> > don't understand the racy code path. The last google problem I saw had
> > s_dentry field as NULL.
> 
> Please take a look at the following message.
> 
>   http://article.gmane.org/gmane.linux.kernel/521729
> 
> I could reproduce both races on my test machine fairly reliable with
> parallel find, cat, mount/mount while repeatedly ins/rmmoding a libata
> driver.
> 
Thanks for the pointer.. earlier it got buried in the fat rework..


> >>> This is an interim solution for -stable kernels.  sysfs reclamation is
> >>> disabled by default and can be enabled by using sysfs.enable_reclaim
> >>> kernel parameter.  Note that dentries are still created on demand, so
> >>> attribute and symlinks nodes aren't allocated on creation.  They're
> >>> allocated on first lookup and deallocated when the sysfs node is
> >>> removed.
> >> Ick, this is going to kill memory on big boxes (s390 and others) and I
> >> don't really want to apply this it if at all possible.
> >>
> > At least not make it default. This might create boot issues with these
> > boxes. 
> 
> Which makes oopsing the default.  Fun!  :-)
> 
but.. avoid oops by not booting at all is more fun !! ;-)


> >> Maneesh, any other thoughts?
> >>
> > I actually wanted to investigate this oops but left it considering the
> > rework being done by Tejun. If this still make sense we can have some
> > more debug code stuffed there or get a crashdump (kdump) to get better
> > understanding of the race.
> 
> The above message contains analysis of both races.  I just ported the
> fixes.  I have a different test machine now and can't reproduce the
> races with this one yet so I couldn't verify whether the patches
> actually fix the problem.  I'll post the patches anyway.  If anyone can
> reproduce these races, please verify the posted patches fix the problem.
> 

I would prefer fixing the race instead of making attributes non-reclaimable.

Thanks
Maneesh



-- 
Maneesh Soni
Linux Technology Center,
IBM India Systems and Technology Lab, 
Bangalore, India