Re: Files full of zeros with coreutils-8.11 and xfs (FIEMAP related?)

[Dave Chinner]

On Sat, Apr 16, 2011 at 02:05:51PM +0800, Yongqiang Yang wrote:
> On Sat, Apr 16, 2011 at 8:50 AM, Dave Chinner <address@hidden> wrote:
> > On Thu, Apr 14, 2011 at 11:01:04PM -0600, Andreas Dilger wrote:
> >> On 2011-04-14, at 6:09 PM, Dave Chinner <address@hidden>
> >> wrote:
> >> > No, this was explicitly laid out in the fiemap interface
> >> > discussions - it's up to the applicaiton to decide if it needs
> >> > to do a sync first. That's what the FIEMAP_FLAG_SYNC control
> >> > flag is for.  This forces the fiemap call to do a fsync _before_
> >> > getting the mapping. If you want to know the exact layout of the
> >> > file is, then you must use this flag.
> >> >
> >> > Even so, it is recognised that this is racy - any use of the
> >> > block map has a time-of-read-to-time-of-use race condition that
> >> > means you have to _verify_ the copy after it completes. FYI,
> >> > that's what xfs_fsr does when copying based on extent maps - if
> >> > the inode has changed in _any way_ during the copy, it aborts
> >> > the copy of that file.
> >> >
> >> > i.e. using fiemap for copying is at best a *hint* about the
> >> > regions that need copying, and it is in no way a guarantee that
> >> > you'll get all the information you need to make accurate copy
> >> > even if you do use the synchronous variant.
> >>
> >> I would tend to agree with Pádraig. If there is data in the
> >> mapping (regardless of whether it is on disk or not), the FIEMAP
> >> should return this to the caller.  The SYNC flag is only intended
> >> to flush the data to disk for tools that are doing
> >> direct-to-disk operations on the data.
> >
> > What you are suggesting is that FIEMAP needs to be page cache
> > coherent, and that is far, far away from the intended use of the
> > interface. Even consiering that you need to looking for active pages
> > in the page cache when mapping extents say to me that you are
> > doing something very wrong.
> >
> > Unwritten extents remain unwritten until the data is physically
> > written to them. Therefore, to change their state, you need to sync
> No, buffered writes change their state without sync.

They shouldn't.

> > the data covering the range.  _Lying_ about whether an extent is in
> > the unwritten state is a really bad precedence to set, especially as
> > it is then guaranteed to change state when a crash occurs (Why did
> > recovery zero out my file? FIEMAP said it contained data before my
> > system crashed!).
> 
> All filesystems have metadata in memory which is not flushed to
> permanent storage. e.g. if a extent exists in memory, but itself and
> corresponding data are not flushed to permanent storage.

Sure, but in the case of unwritten extents, XFS does not change the
metadata state in memory until *after the physical IO is completed*.
I'm pretty sure that btrfs is the same.

IOWs, despite the fact that a buffered write has occurred, no
metadata has changed state in memory, and the extents are still
unwritten in both memory and on disk....

Cheers,

Dave.
-- 
Dave Chinner
address@hidden