On Wed, 26 Feb 2014 12:34:26 +1100 Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> On Tue, Feb 25, 2014 at 03:41:28PM -0800, Andrew Morton wrote:
> > On Tue, 25 Feb 2014 15:23:35 -0800 (PST) Hugh Dickins <hughd@xxxxxxxxxx>
> > wrote:
> > > On Tue, 25 Feb 2014, Dave Chinner wrote:
> > > > On Tue, Feb 25, 2014 at 02:16:01PM +1100, Stephen Rothwell wrote:
> > > > > On Mon, 24 Feb 2014 11:57:10 +1100 Dave Chinner <david@xxxxxxxxxxxxx>
> > > > > wrote:
> > > FALLOC_FL_COLLAPSE_RANGE: I'm a little sad at the name COLLAPSE,
> > > but probably seven months too late to object. It surprises me that
> > > you're doing all this work to deflate a part of the file, without
> > > the obvious complementary work to inflate it - presumably all those
> > > advertisers whose ads you're cutting out, will come back to us soon
> > > to ask for inflation, so that they have somewhere to reinsert them ;)
> > Yes, I was wondering that. Why not simply "move these blocks from here
> > to there".
> And open a completely unnecessary can of worms to do with
> behavioural and implementation corner cases?
But it's general.
> Do you allow it to destroy data by default? Or only allow moves into
> What do you do with range the data is moved out of? Does it just
> become a hole? What happens if the range overlaps EOF - does that
> change the file size?
> What if you want to move the range beyond EOF?
> What if the source and destination ranges overlap?
Don't screw it up.
> What happens when you move the block at EOF into the middle of a
> file - do you end up with zeros padding the block and the file size
> having to be adjusted accordingly? Or do we have to *copy* all the
> data in high blocks down to fill the hole in the block?
I don't understand that. Move the block(s) and truncate to the new
> What behaviour should we expect if the filesystem can't implement
> the entire move atomically and we crash in the middle of the move?
What does collapse_range do now?
If it's a journaled filesystem, it shouldn't screw up. If it isn't, fsck.
> I can keep going, but I'll stop here - you get the idea.
None of this seems like rocket science.
> In comparison, collapse range as a file data manipulation has very
> specific requirements and from that we can define a simple, specific
> API that allows filesystems to accelerate that operation by extent
> manipulation rather than read/memcpy/write that the applications are
> currently doing for this operation.... IOWs, collapse range is a
> simple operation, "move arbitrary blocks from here to there" is a
> nightmare both from the specification and the implementation points
> of view.
collapse_range seems weird, arbitrary and half-assed. "Why didn't they
go all the way and do it properly".