Re: 2.6.2-mm1 aka "Geriatric Wombat" DIO read race still fails

[Daniel McNeil]

On Thu, 2004-02-05 at 16:07, Andrew Morton wrote:
> Daniel McNeil <daniel@xxxxxxxx> wrote:
> >
> > Andrew,
> > 
> > I tested 2.6.2-mm1 on an 8-proc running 6 copies of the read_under
> > test and all 6 read_under tests saw uninitialized data in less than 5
> > minutes. :(
> 
> The performance implications of synchronising behind kjournald writes for
> normal non-blocking writeback are bad.  Can you detail what you now think
> is the failure mechanism?
> 

I think the problem is that any block_write_full_page(WB_SYNC_NONE)
that hits a page that has a buffer in process of being written will
get PageWriteback cleared even though the i/o has not completed.
(The buffer will be locked, but buffer_dirty() is cleared, so
 __block_write_full_page() will SetPageWriteback(); unlock_page();
 see no buffer were submitted and call end_page_writeback())

Any subsequent filemap_write_and_wait() or filemap_fdatawrite() /
filemap_fdatawait will never wait for that i/o.  So this could
potentially be a problem for more than just DIO.

BTW: 2.4 __block_write_full_page() always did a lock_buffer(), so
it waits for i/o in flight.

I agree though, it would be best if non-sync __block_write_full_page()
would not block on buffers in flight.  Somehow we need to move the
clearing of PageWriteback() until after the buffer has been written
even for the case where ll_rw_block() is called.

Thoughts?

Daniel


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/