On Fri, 15 Aug 2008 18:09:58 -0400
Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote:
> On Fri, Aug 15, 2008 at 02:08:50PM +1000, Lachlan McIlroy wrote:
> > With the help from some tracing I found that we try to map extents beyond
> > eof when doing a direct I/O read. It appears that the way to inform the
> > generic direct I/O path (ie do_direct_IO()) that we have breached eof is
> > to return an unmapped buffer from xfs_get_blocks_direct(). This will cause
> > do_direct_IO() to jump to the hole handling code where is will check for
> > eof and then abort.
> > This problem was found because a direct I/O read was trying to map beyond
> > eof and was encountering delayed allocations. The delayed allocations
> > beyond
> > eof are speculative allocations and they didn't get converted when the
> > direct
> > I/O flushed the file because there was only enough space in the current AG
> > to convert and write out the dirty pages within eof. Note that
> > xfs_iomap_write_allocate() wont necessarily convert all the delayed
> > allocation
> > passed to it - it will return after allocating the first extent - so if the
> > delayed allocation extends beyond eof then it will stay that way.
> > This change will detect a direct I/O read beyond eof:
> The change looks good to me, but I really think the direct I/O could
> should never send down requests like this down to the filesystems. akpm
> and -fsdevel Cc'ed.
Oh gee, I forget, and so many people have done drivebys on that code...
We _could_ add additional i_size checking into direct-io.c but bear in
mind that it would be best-effort unreliable stuff. The code will
still be tripped up by concurrent extends and concurrent truncates.
So we'll still end up calling the fs for blocks outside i_size, only
less commonly. I think.