On Thu, Feb 18, 2010 at 08:13:55AM +1100, Dave Chinner wrote:
> On Wed, Feb 17, 2010 at 03:31:06AM -0500, Christoph Hellwig wrote:
> > On Wed, Feb 17, 2010 at 02:55:11PM +1100, Dave Chinner wrote:
> > > You've changed a local variable "pos" which had the value
> > > iocb->ki_pos to a function parameter of the same name which has a
> > > different value. Given this, all the existing uses of "pos" in this
> > > function need to be converted to "iocb->ki_pos" as the old
> > > xfs_write() never saw the original "pos" variable passed to
> > > xfs_file_aio_write().
> > Oh, I should have explained this in more detail. The aio_read/aio_write
> > ABI both has a pos argument and the file position in iocb->ki_pos.
> > They were added for allowing aio that does partial I/O in each method
> > call using retries, but we don't actually use the anywhere. Thus these
> > two are always these same and we even enforce that with a
> > BUG_ON(iocb->ki_pos != pos);
> > in the code (both old and new). Now how does the old pos local variable
> > come in play? The old code didn't want to pass the kiocb to the
> > low-level xfs-write function, but as want the offset it passes a pointer
> > to iocb->ki_pos, which is called offset. We take a local copy of it
> > before we might start modifying it, which we call pos. pos gets updated
> > early in generic_write_checks if this is an O_APPEND write, but
> > otherwise stays immutable and marks the position where this write
> > started, while iocb->ki_pos (aka the old offset) gets updated by
> > generic_file_direct_write / generic_file_buffered_write to the new
> > file position after the I/O was done.
> Ok, my misunderstanding. I'll go back and review it again with this
> in mind.
It looks ok having taken this into account. I think I originally
read the BUG_ON(iocb->ki_pos != pos) as meaning the two values were
not equivalent because that is what an ASSERT(iocb->ki_pos != pos)
would mean. Anyway:
Reviewed-by: Dave Chinner <david@xxxxxxxxxxxxx>