Re: [PATCH 3/3] block: implement (some of) fallocate for block devices

[Mike Snitzer]

On Mon, Mar 21 2016 at  3:11pm -0400,
Darrick J. Wong <darrick.wong@xxxxxxxxxx> wrote:

> On Mon, Mar 21, 2016 at 02:52:00PM -0400, Mike Snitzer wrote:
> > On Tue, Mar 15, 2016 at 3:42 PM, Darrick J. Wong
> > <darrick.wong@xxxxxxxxxx> wrote:
> > > After much discussion, it seems that the fallocate feature flag
> > > FALLOC_FL_ZERO_RANGE maps nicely to SCSI WRITE SAME; and the feature
> > > FALLOC_FL_PUNCH_HOLE maps nicely to the devices that have been
> > > whitelisted for zeroing SCSI UNMAP.  Punch still requires that
> > > FALLOC_FL_KEEP_SIZE is set.  A length that goes past the end of the
> > > device will be clamped to the device size if KEEP_SIZE is set; or will
> > > return -EINVAL if not.  Both start and length must be aligned to the
> > > device's logical block size.
> > >
> > > Since the semantics of fallocate are fairly well established already,
> > > wire up the two pieces.  The other fallocate variants (collapse range,
> > > insert range, and allocate blocks) are not supported.
> > 
> > I'd like to see fallocate (block allocation) extend down to DM thinp.
> > This more traditional use of fallocate would be useful for ensuring
> > ENOSPC won't occur -- especially important if the FS has committed
> > space in response to fallocate.  As of now fallocate doesn't inform DM
> > thinp at all.  Curious why you decided not to wire it up?
> 
> I don't know what to wire it up to. :)

Fair enough.  Yes something needs to be invented.
 
> I didn't find any blkdev_* function that looked encouraging, though I
> haven't dug too deeply into bfoster's "prototype a block reservation
> allocation model" patchset yet.  At a high level I'd guess that would
> be a reasonable piece to connect to?  It looks like the piece I want
> is blk_provision_space().

Yes, something like that.

> > But I'm not sure what "it" (the "allocate blocks" variant) even is
> > given falloc.h doesn't show anything like "_ALLOCATE_BLOCKS"...
> 
> The default behavior of fallocate is to allocate blocks, which means
> that one invokes it by not passing any mode flags (except possibly
> KEEP_SIZE).

OK.

> > It would require a new block interface to pass the fallocate extent
> > down.  But it seems bizarre to implement "some of" fallocate but not
> > the most widely used case for fallocate.
> 
> Agreed.  I'd like to get the existing functionality wired up sooner than
> later, and plumbing "allocate blocks" down to thinp can be done as a
> followup.
> 
> (Or stall long enough that it becomes one patchset.)

Sure, sounds good.  Glad we're in agreement.