Re: O_DIRECT broken in 2.5.3-preX ?

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Jan 24, 2002 at 01:52:04PM -0800, Badari Pulavarty wrote:
> Hi,
> 
> I am reading the O_DIRECT code patch for 2.5.3-pre4. I was wondering
> how is this working in 2.5.X ? Here is my concern:
> 
> generic_direct_IO() creates a blocks[] list and passes it to
> brw_kiovec() with a single kiobuf.
> 	
> 	retval = brw_kiovec(rw, 1, &iobuf, inode->i_dev, blocks, blocksize);
> 
> But brw_kiovec() uses only b[0] to call ll_rw_bio().
> 
> 	for (i = 0; i < nr; i++) {
>                 iobuf = iovec[i];
>                 iobuf->errno = 0;
> 
>                 ll_rw_kio(rw, iobuf, dev, b[i] * (size >> 9));
>         }
> 
> 
> Note that nr = 1 here. ll_rw_kio() uses b[0] as starting sector
> and does the entire IO (for iobuf->length). This is wrong !!!
> It is doing IO from wrong blocks.  Some one should use other 
> block numbers from blocks[] list. Isn't it ?

correct. It seems like somebody broke the semantics of the block field
of brw_kiovec without updating the callers, and furthmore now it is
impossible to use brw_kiovec for O_DIRECT because it is not going to
submit physically contigous I/O.

Andrea