Re: [PATCH 3/3] xfs: measure all contiguous previous extents for prealloc size

[Brian Foster <bfoster@redhat.com>]

On Wed, May 20, 2020 at 12:48:53PM -0700, Darrick J. Wong wrote:
> On Wed, May 20, 2020 at 09:23:34AM -0400, Brian Foster wrote:
> > On Tue, May 19, 2020 at 08:48:27AM -0400, Brian Foster wrote:
> > > On Mon, May 18, 2020 at 05:49:23PM -0700, Darrick J. Wong wrote:
> > > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > > 
> > > > When we're estimating a new speculative preallocation length for an
> > > > extending write, we should walk backwards through the extent list to
> > > > determine the number of number of blocks that are physically and
> > > > logically contiguous with the write offset, and use that as an input to
> > > > the preallocation size computation.
> > > > 
> > > > This way, preallocation length is truly measured by the effectiveness of
> > > > the allocator in giving us contiguous allocations without being
> > > > influenced by the state of a given extent.  This fixes both the problem
> > > > where ZERO_RANGE within an EOF can reduce preallocation, and prevents
> > > > the unnecessary shrinkage of preallocation when delalloc extents are
> > > > turned into unwritten extents.
> > > > 
> > > > This was found as a regression in xfs/014 after changing delalloc writes
> > > > to create unwritten extents during writeback.
> > > > 
> > > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > > > ---
> > > >  fs/xfs/xfs_iomap.c |   63 +++++++++++++++++++++++++++++++++++++++++++---------
> > > >  1 file changed, 52 insertions(+), 11 deletions(-)
> > > > 
> > > > 
> > > > diff --git a/fs/xfs/xfs_iomap.c b/fs/xfs/xfs_iomap.c
> > > > index ac970b13b1f8..2dffd56a433c 100644
> > > > --- a/fs/xfs/xfs_iomap.c
> > > > +++ b/fs/xfs/xfs_iomap.c
> > ...
> > > > @@ -413,15 +453,16 @@ xfs_iomap_prealloc_size(
> > > >  	 * preallocation size.
> > > >  	 *
> > > >  	 * If the extent is a hole, then preallocation is essentially disabled.
> > > > -	 * Otherwise we take the size of the preceding data extent as the basis
> > > > -	 * for the preallocation size. If the size of the extent is greater than
> > > > -	 * half the maximum extent length, then use the current offset as the
> > > > -	 * basis. This ensures that for large files the preallocation size
> > > > -	 * always extends to MAXEXTLEN rather than falling short due to things
> > > > -	 * like stripe unit/width alignment of real extents.
> > > > +	 * Otherwise we take the size of the contiguous preceding data extents
> > > > +	 * as the basis for the preallocation size. If the size of the extent
> > > 
> > > I'd refer to it as the "size of the contiguous range" or some such since
> > > we're now talking about aggregating many extents to form the prealloc
> > > basis.
> > > 
> > > I am a little curious if there's any noticeable impact from having to
> > > perform the worst case extent walk in this path. For example, suppose we
> > > had a speculatively preallocated file where we started writing (i.e.
> > > converting) every other unwritten block such that this path had to walk
> > > an extent per block until hitting the 8g max (8g == 2097152 4k blocks).
> > > I guess the hope is that either most of those blocks wouldn't have been
> > > written back and converted by the time we'd trigger the next post-eof
> > > prealloc or that it would just be a wash in the stream of staggered
> > > writes falling into our max sized preallocations. Either way, I think
> > > it's more important to maintain the existing heuristic so this seems
> > > reasonable from that perspective.
> > > 
> > 
> > I had a system spinning yesterday to create this worst case condition
> > across a couple max sized extents. Testing out the preallocation path, I
> > see a hit from ~5ms to ~35ms between the baseline variant and the
> > updated calculation algorithm. Note that the baseline here is only doing
> > a 16 block prealloc vs. 8gb with the fix, but regardless a 30ms
> > difference for an 8gb allocation doesn't really seem noticeable enough
> > to matter. IOW, I think we can disgregard this particular concern...
> 
> <nod> I'd figured that walking backwards through the incore extent tree
> shouldn't be all that costly, particularly since it tends to result in
> more aggressive speculative preallocation.
> 

Yeah, I was more curious about going from oneshot "check the prev
extent" logic to "walk N extents" logic with a significantly higher
upper bound and wanted to make sure we've analyzed worst case behavior.

> > > This scenario also makes me wonder if we should consider continuing to
> > > do write time file size extension zeroing in certain cases vs. leaving
> > > around unwritten extents. For example, the current post-eof prealloc
> > > behavior is that writes that jump over current EOF will zero (via
> > > buffered writes) any allocated blocks (delalloc or physical) between
> > > current EOF and the start of the write.
> 
> Right...
> 
> > > That behavior doesn't change with delalloc -> unwritten if the
> > > prealloc is still delalloc at the time of the extending write, but
> > > we'd presumably skip those zeroing writes if the prealloc had been
> > > converted to real+unwritten first.
> 
> ...though I wonder how often we'll encounter the situation where we've
> created a posteof preallocation, and we end up with written extents
> beyond EOF, and then the next write jumps past EOF?
> 

I don't think it's all that uncommon. The first contributing factor is
preallocation size. If a write workload is sparsely appending to a file
and sees one good speculative preallocation condition, further extending
writes are more likely to fall into the preallocation and essentially
compound the behavior for subsequent preallocations. The second
contributing factor is writeback converting the delalloc extent. I guess
we have less control over that between writeback heuristics, workload
(fsync?), memory availability, etc., but suffice it to say that the
larger the file and EOF extent grows, the more likely writeback converts
the extent to real (now unwritten) blocks.

> > > Hmm.. that does
> > > seem a little random to me, particularly if somebody starts to see
> > > unwritten extents sprinkled throughout a file that never explicitly saw
> > > preallocation. Perhaps we should avoid converting post-eof blocks? OTOH,
> > > unwritten probably makes sense for large jumps in EOF vs zeroing GBs of
> > > blocks, so one could argue that we should convert such ranges (if still
> > > delalloc) rather than zero them at all. Thoughts? We should probably
> > > work this out one way or another before landing the delalloc ->
> > > unwritten behavior..
> 
> /me wonders if that will be a problem in practice?  If writeback starts,
> it'll (try to) convert the delalloc reservation into an unwritten extent
> (even if the extent crosses EOF).  Writeback completion will convert the
> written parts to regular extents, leaving the rest unwritten.
> 
> iomap_zero_range is a NOP on unwritten extents, so starting a new write
> far beyond EOF will do very little work to make sure [oldsize, write
> start] range is zero.
> 
> (Or am I misunderstanding this...?)
> 

Nope, that's what happens. It's not a correctness issue. I'm just
calling out the subtle behavior change in that we can potentially leave
unwritten post-eof ranges sprinkled around as unwritten extents and
asking whether we think that might cause any notable issues for users.

We used to have a steady trickle of questions along the lines of "why
does my file/fs use so much space?" that boiled down to speculative
preallocation. The solution was usually more user education (i.e., FAQ
updates) and I think users have kind of become used to it by now, but
the broader point is that some users might notice such things and wonder
why some workload that previously created large files with fewer large
extents now creates files with hundreds (or more) extents purely due to
extent state of unwritten ranges.

I'm not totally convinced it matters. It could very well be just another
case of more education, but I wanted to at least make sure we discussed
it. ;) What particularly annoys me about this is that the resulting
behavior is so inconsistent. Some workloads that might involve fsync or
otherwise more frequent writeback events could see more unwritten
extents whereas others with very similar write patterns might see none
at all. Conversely, some users might continue to see huge post-eof
zeroing writes while others won't depending on whether a post-eof write
beats out writeback of the associated of extent.

> > Thinking about this one a little more.. it seems it's probably not worth
> > conditionally changing post-eof conversion behavior. Since there are
> > cases where we probably want post-eof prealloc to remain as unwritten if
> > it's carried within i_size, that would either be extra code for post-eof
> > blocks or would split up any kind of heuristic for manually zeroing such
> > blocks. ISTM that the right approach might be to convert all delalloc ->
> 
> I might just leave it all unwritten and not try to be clever about
> pre-zeroing when we extend EOF because that just adds more complexity.
> 

Well we do already have the post-eof zeroing logic, but indeed it would
be more complexity to conditionalize it on something other than extent
state. I guess we'd need to pass a size hint or something to iomap or
otherwise modify our callback path to detect this situation. There's
also the question of whether to overwrite explicit (fallocate) post-eof
preallocation or not...

> > unwritten (as this series already does), then perhaps consider a write
> > time heuristic that would perform manual zeroing of extents if certain
> > conditions are met (e.g., for unwritten extents under a certain size).
> 
> ISTR that ext4 has some sort of heuristic to do that.  I'm not sure
> which is better as writes get relatively more expensive, but so long as
> the amount of zeroing is less than an erase block size it probably
> doesn't matter...
> 

Yeah, though note again that historical behavior is to perform buffered
writes to zero the range regardless of its size or delalloc vs. real
allocation state (explicit preallocation notwithstanding). We could use
a cutoff size of tens of MBs or more and still be pretty conservative
from a behavioral change standpoint.

> > The remaining question to me is whether we should include something like
> > that before changing delalloc behavior or consider it separately as an
> > optimization...
> 
> Separately.  It's getting late in the cycle. ;)
> 

To clarify, my question is whether we should require manual zeroing
logic for some cases before changing delalloc allocation behavior or if
we're Ok with whatever the "worst case" behavior is (and thus consider
such manual zeroing a fragmentation mitigation optimization that can
come later). I think this is independent of wherever the current release
cycle is at because it's more a question of getting it right vs. getting
it done. :)

Brian

> > Brian
> > 
> > P.S., BTW I forgot to mention on my last pass of this series that it
> > probably makes sense to reorder these behavioral fixup patches to
> > precede the patch that actually changes delalloc to allocate unwritten
> > extents.
> 
> Yeah, I'll do that.  I think I mostly like Christoph's rewrite of the
> third patch of the series.
> 
> --D
> 
> > > Brian
> > > 
> > > > +	 * is greater than half the maximum extent length, then use the current
> > > > +	 * offset as the basis. This ensures that for large files the
> > > > +	 * preallocation size always extends to MAXEXTLEN rather than falling
> > > > +	 * short due to things like stripe unit/width alignment of real
> > > > +	 * extents.
> > > >  	 */
> > > > -	if (prev.br_blockcount <= (MAXEXTLEN >> 1))
> > > > -		alloc_blocks = prev.br_blockcount << 1;
> > > > +	if (plen <= (MAXEXTLEN >> 1))
> > > > +		alloc_blocks = plen << 1;
> > > >  	else
> > > >  		alloc_blocks = XFS_B_TO_FSB(mp, offset);
> > > >  	if (!alloc_blocks)
> > > > 
> > > 
> > 
>