On Mon, Mar 17, 2014 at 01:28:04AM +0000, Al Viro wrote:
> On Mon, Mar 17, 2014 at 12:29:18AM +0000, Al Viro wrote:
> > I think I know what's going on - O_DIRECT write starting a bit before
> > EOF on a file with the last extent that can be grown. It fills
> > a buffer_head with b_size extending quite a bit past the EOF; the
> > blocks are really allocated. What causes the problem is that we
> > have the flags set for the *first* block. IOW, buffer_new(bh) is
> > false - the first block has already been allocated. And for
> > direct-io.c it means "no zeroing the tail of the last block".
> BTW, that's something I have directly observed - xfs_get_blocks_direct()
> called with iblock corresponding to a bit under 16Kb below EOF and
> returning with ->b_size equal to 700K and ->b_flags not containing BH_New.
What's the userspace IO pattern that triggers this?
> IOW, we really can't mix new and old blocks in that interface - not enough
> information is passed back to caller to be able to decide what does and
> what does not need zeroing out. It should be either all-new or all-old.
Right, and XFS should not be mixing old and new in the way you are
describing, and that's what I can't reproduce. See my reply on the
other thread. Probably best to continue there...
> And it's not just the EOF, of course - the beginning of a hole in a sparse
> file isn't any different from the end of file in that respect.
Except that XFS treats that differently - it does allocation as
unwritten extents there, and any mapping that covers an unwritten
block will always result in buffer_new() getting set...