Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1206753

Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock

Show all headers | View raw

On Wed 12-08-15 07:48:22, Dave Chinner wrote:
> On Tue, Aug 11, 2015 at 04:51:22PM +0000, Wilcox, Matthew R wrote:
> > The race that you're not seeing is page fault vs page fault.  Two
> > threads each attempt to store a byte to different locations on the
> > same page.  With a read-mutex to exclude truncates, each thread
> > calls ->get_block.  One of the threads gets back a buffer marked
> > as BH_New and calls memset() to clear the page.  The other thread
> > gets back a buffer which isn't marked as BH_New and simply inserts
> > the mapping, returning to userspace, which stores the byte ...
> > just in time for the other thread's memset() to write a zero over
> > the top of it.
> 
> So, this is not a truncate race that the XFS MMAPLOCK solves.
> 
> However, that doesn't mean that the DAX code needs to add locking to
> solve it. The race here is caused by block initialisation being
> unserialised after a ->get_block call allocates the block (which the
> filesystem serialises via internal locking). Hence two simultaneous
> ->get_block calls to the same block is guaranteed to have the DAX
> block initialisation race with the second ->get_block call that says
> the block is already allocated.
> 
> IOWs, the way to handle this is to have the ->get_block call handle
> the block zeroing for new blocks instead of doing it after the fact
> in the generic DAX code where there is no fine-grained serialisation
> object available. By calling dax_clear_blocks() in the ->get_block
> callback, the filesystem can ensure that the second racing call will
> only make progress once the block has been fully initialised by the
> first call.
> 
> IMO the fix is - again - to move the functionality into the
> filesystem where we already have the necessary exclusion in place to
> avoid this race condition entirely.

I'm somewhat sad to add even more functionality into the already loaded
block mapping interface - we can already allocate delalloc blocks, unwritten
blocks, uninitialized blocks, and now also pre-zeroed blocks. But I agree
fs already synchronizes block allocation for a given inode so adding the
pre-zeroing there is pretty easy. Also getting rid of unwritten extent
handling from DAX code is a nice bonus so all in all I'm for this approach.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Back to linux.kernel | Previous | Next — Previous in thread | Find similar | Unroll thread

Thread

[PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> - 2015-08-10 17:20 +0200
  Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Jan Kara <jack@suse.cz> - 2015-08-11 10:20 +0200
    Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Dave Chinner <david@fromorbit.com> - 2015-08-11 11:40 +0200
      Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-11 13:20 +0200
      Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock "Kirill A. Shutemov" <kirill@shutemov.name> - 2015-08-11 14:10 +0200
      Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Jan Kara <jack@suse.cz> - 2015-08-11 16:00 +0200
        Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-11 16:40 +0200
          Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock "Kirill A. Shutemov" <kirill@shutemov.name> - 2015-08-11 17:30 +0200
            Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-11 18:20 +0200
              Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock "Kirill A. Shutemov" <kirill@shutemov.name> - 2015-08-11 22:30 +0200
                Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-12 10:00 +0200
          RE: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock "Wilcox, Matthew R" <matthew.r.wilcox@intel.com> - 2015-08-11 19:00 +0200
            Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-11 20:50 +0200
            Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Dave Chinner <david@fromorbit.com> - 2015-08-11 23:50 +0200
              Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Boaz Harrosh <boaz@plexistor.com> - 2015-08-12 11:00 +0200
              Re: [PATCH, RFC 2/2] dax: use range_lock instead of i_mmap_lock Jan Kara <jack@suse.cz> - 2015-08-13 13:40 +0200

csiph-web