Re: [PATCH 14/15] xfs: compute absolute maximum nlevels for each btree type

[Dave Chinner <david@fromorbit.com>]

On Wed, Oct 13, 2021 at 02:36:33PM -0700, Darrick J. Wong wrote:
> On Wed, Oct 13, 2021 at 06:57:43PM +1100, Dave Chinner wrote:
> > On Tue, Oct 12, 2021 at 04:33:50PM -0700, Darrick J. Wong wrote:
> > > --- a/fs/xfs/libxfs/xfs_alloc_btree.c
> > > +++ b/fs/xfs/libxfs/xfs_alloc_btree.c
> > > @@ -582,6 +582,19 @@ xfs_allocbt_maxrecs(
> > >  	return blocklen / (sizeof(xfs_alloc_key_t) + sizeof(xfs_alloc_ptr_t));
> > >  }
> > >  
> > > +/* Compute the max possible height of the maximally sized free space btree. */
> > > +unsigned int
> > > +xfs_allocbt_absolute_maxlevels(void)
> > > +{
> > > +	unsigned int		minrecs[2];
> > > +
> > > +	xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_alloc_rec_t),
> > > +			sizeof(xfs_alloc_key_t) + sizeof(xfs_alloc_ptr_t));
> > > +
> > > +	return xfs_btree_compute_maxlevels(minrecs,
> > > +			(XFS_MAX_AG_BLOCKS + 1) / 2);
> > > +}
> > 
> > Hmmmm. This is kinds messy. I'd prefer we share code with the
> > xfs_allocbt_maxrecs() function that do this. Not sure "absolute" is
> > the right word, either. It's more a function of the on-disk format
> > maximum, not an "absolute" thing.
> 
> <nod> I'm not passionate about the name one way or the other.
> 
> > I mean, we know that the worst case is going to be for each btree
> > type - we don't need to pass in XFS_BTREE_CRC_BLOCKS or
> > XFS_BTREE_LONG_PTRS to generic code for it to branch multiple times
> > to be generic.
> 
> Yeah, that function was a conditional mess.  I like...
> 
> > Instead:
> > 
> > static inline int
> > xfs_allocbt_block_maxrecs(
> >         int                     blocklen,
> >         int                     leaf)
> > {
> >         if (leaf)
> >                 return blocklen / sizeof(xfs_alloc_rec_t);
> >         return blocklen / (sizeof(xfs_alloc_key_t) + sizeof(xfs_alloc_ptr_t));
> > }
> > 
> > /*
> >  * Calculate number of records in an alloc btree block.
> >  */
> > int
> > xfs_allocbt_maxrecs(
> >         struct xfs_mount        *mp,
> >         int                     blocklen,
> >         int                     leaf)
> > {
> >         blocklen -= XFS_ALLOC_BLOCK_LEN(mp);
> > 	return xfs_allobt_block_maxrecs(blocklen, leaf);
> > }
> > 
> > xfs_allocbt_maxlevels_ondisk()
> > {
> > 	unsigned int		minrecs[2];
> > 
> > 	minrecs[0] = xfs_allocbt_block_maxrecs(
> > 			XFS_MIN_BLOCKSIZE - XFS_BTREE_SBLOCK_LEN, true) / 2;
> > 	minrecs[1] = xfs_allocbt_block_maxrecs(
> > 			XFS_MIN_BLOCKSIZE - XFS_BTREE_SBLOCK_LEN, false) / 2;
> 
> ...this a lot better since one doesn't have to switch back and forth
> between source files to figure out how the computation works.
> 
> However, I want to propose a possibly pedantic addition to the blocksize
> computation for btrees.  We want to compute the maximum btree height
> that we're ever going to see, which means that we are modeling a btree
> with the minimum possible fanout factor.  That means the smallest btree
> nodes possible, and half full.
> 
> min V5 blocksize: 1024 bytes
> V5 btree short header: 56 bytes
> min V5 btree record area: 968 bytes
> 
> min V4 blocksize: 512 bytes
> V4 btree short header: 16 bytes
> min V4 btree record area: 496 bytes
> 
> In other words, the bit above for the allocbt ought to be:
> 
> 	blocklen = min(XFS_MIN_BLOCKSIZE - XFS_BTREE_SBLOCK_LEN,
> 		       XFS_MIN_CRC_BLOCKSIZE - XFS_BTREE_SBLOCK_CRC_LEN);
> 
> Which is very pedantic, since the whole expression /always/ evalulates
> to 496.  IIRC the kernel has enough macro soup to resolve that into a
> constant expression so it shouldn't cost us anything.

Yup, good idea, I'm happy with that - now the code documents the
on-disk format calculation exactly in a single location. :)

> > > --- a/fs/xfs/libxfs/xfs_ialloc.c
> > > +++ b/fs/xfs/libxfs/xfs_ialloc.c
> > > @@ -2793,6 +2793,7 @@ xfs_ialloc_setup_geometry(
> > >  	inodes = (1LL << XFS_INO_AGINO_BITS(mp)) >> XFS_INODES_PER_CHUNK_LOG;
> > >  	igeo->inobt_maxlevels = xfs_btree_compute_maxlevels(igeo->inobt_mnr,
> > >  			inodes);
> > > +	ASSERT(igeo->inobt_maxlevels <= xfs_inobt_absolute_maxlevels());
> > >  
> > >  	/*
> > >  	 * Set the maximum inode count for this filesystem, being careful not
> > > diff --git a/fs/xfs/libxfs/xfs_ialloc_btree.c b/fs/xfs/libxfs/xfs_ialloc_btree.c
> > > index 3a5a24648b87..2e3dd1d798bd 100644
> > > --- a/fs/xfs/libxfs/xfs_ialloc_btree.c
> > > +++ b/fs/xfs/libxfs/xfs_ialloc_btree.c
> > > @@ -542,6 +542,25 @@ xfs_inobt_maxrecs(
> > >  	return blocklen / (sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t));
> > >  }
> > >  
> > > +/* Compute the max possible height of the maximally sized inode btree. */
> > > +unsigned int
> > > +xfs_inobt_absolute_maxlevels(void)
> > > +{
> > > +	unsigned int		minrecs[2];
> > > +	unsigned long long	max_ag_inodes;
> > > +
> > > +	/*
> > > +	 * For the absolute maximum, pretend that we can fill an entire AG
> > > +	 * completely full of inodes except for the AG headers.
> > > +	 */
> > > +	max_ag_inodes = (XFS_MAX_AG_BYTES - (4 * BBSIZE)) / XFS_DINODE_MIN_SIZE;
> > > +
> > > +	xfs_btree_absolute_minrecs(minrecs, 0, sizeof(xfs_inobt_rec_t),
> > > +			sizeof(xfs_inobt_key_t) + sizeof(xfs_inobt_ptr_t));
> > > +
> > > +	return xfs_btree_compute_maxlevels(minrecs, max_ag_inodes);
> > > +}
> > 
> > We've got two different inobt max levels on disk. The inobt which has v4
> > limits, whilst the finobt that has v5 limits...
> 
> <nod> I'll make it return the larger of the two heights, though the
> inode btree is always going to win due to its smaller minimum block size.

Yup, I expect so, but it would be good to make it explicit :)

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com