Re: [PATCH] hugetlbfs: lockdep annotate root inode properly

["Aneesh Kumar K.V" <aneesh.kumar@linux.vnet.ibm.com>]

On Thu, 8 Mar 2012 13:40:50 -0800, Andrew Morton <akpm@linux-foundation.org> wrote:
> On Thu, 8 Mar 2012 15:19:27 -0600
> Tyler Hicks <tyhicks@canonical.com> wrote:
> 
> > > 
> > > 
> > > Sigh.  Was lockdep_annotate_inode_mutex_key() sufficiently
> > > self-explanatory to justify leaving it undocumented?
> > > 
> > > <goes off and reads e096d0c7e2e>
> > > 
> > > OK, the patch looks correct given the explanation in e096d0c7e2e, but
> > > I'd like to understand why it becomes necessary only now.
> > > 
> > > > NOTE: This patch also require 
> > > > http://thread.gmane.org/gmane.linux.file-systems/58795/focus=59565
> > > > to remove the lockdep warning
> > > 
> > > And that patch has been basically ignored.
> > 
> > Al commented on it here:
> > 
> > https://lkml.org/lkml/2012/2/16/518
> > 
> > He said that while my patch is correct, taking i_mutex inside mmap_sem
> > is still wrong.
> 
> OK, thanks, yup.  Taking i_mutex in file_operations.mmap() is wrong.
> 
> Is hugetlbfs actually deadlockable because of this, or is it the case
> that the i_mutex->mmap_sem ordering happens to never happen for this
> filesystem?  Although we shouldn't go and create incompatible lock
> ranking rules for different filesystems!
> 
> So we need to pull the i_mutex out of hugetlbfs_file_mmap().  What's it
> actually trying to do in there?  If we switch to
> i_size_read()/i_size_write() then AFAICT the problem comes down to
> hugetlb_reserve_pages().
> 
> hugetlb_reserve_pages() fiddles with i_mapping->private_list and the fs
> owns private_list and is free to use a lock other than i_mutex to
> protect it.  (In fact i_mapping.private_lock is the usual lock for
> private_list).
> 
> 
> 
> So from a quick scan here I'm thinking that a decent fix is to remove
> the i_mutex locking from hugetlbfs_file_mmap(), switch
> hugetlbfs_file_mmap() to i_size_read/write then use a hugetlb-private
> lock to protect i_mapping->private_list.  region_chg() will do
> GFP_KERNEL allocations under that lock, so some care is needed.
> 

But as per 7762f5a0b709b415fda132258ad37b9f2a1db994 i_size_write should
always happen with i_mutex held 

-aneesh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>