Re: 2.4.20pre5aa1

[Christoph Hellwig <hch@infradead.org>]

On Thu, Sep 05, 2002 at 09:06:00PM +0200, Andrea Arcangeli wrote:
> On Thu, Sep 05, 2002 at 07:48:24PM +0100, Christoph Hellwig wrote:
> > On Thu, Sep 05, 2002 at 08:41:25PM +0200, Andrea Arcangeli wrote:
> > > other fs, if you're not holding the i_sem (and you certainly aren't
> > > holding the i_sem that frequently, you don't even for writes).
> > 
> > Except of O_DIRECT writes we _do_ hold i_sem, btw.
> 
> can't you end with this?
> 
> 					O_DIRECT write
> 					write finishes
> 	truncate drops the write
> 	truncate set i_sem to 0
> 					write set i_sem to something

s/i_sem/i_size/g ?

> and the fs is then corrupt? (very minor corruption of course and
> extremely hard to trigger, trivially solvable by an fsck, ext[23] had
> similar issues too with the get_block failures with < PAGE_SIZE
> softblocksize, fixed around 2.4.19, that was certainly easier to
> reproduce btw)

Won't happen:


					O_DIRECT write starts
					+ takes XFS iolock exclusive
					- invalidates pagecache
					+ downgrades iolock to shared
					- perform write

	xfs_setattr for truncate called
	+ takes XFS iolock shared
	  -> blocks

					- write i_size to something
					+ releases iolock
	  -> gets woken