Re: [PATCH 4/9] xfs_repair: rebuild free space btrees with bulk loader

[Brian Foster <bfoster@redhat.com>]

On Fri, May 29, 2020 at 02:39:37PM -0700, Darrick J. Wong wrote:
> On Thu, May 28, 2020 at 11:10:44AM -0400, Brian Foster wrote:
> > On Tue, May 19, 2020 at 06:51:08PM -0700, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <darrick.wong@oracle.com>
> > > 
> > > Use the btree bulk loading functions to rebuild the free space btrees
> > > and drop the open-coded implementation.
> > > 
> > > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > > ---
> > >  libxfs/libxfs_api_defs.h |    3 
> > >  repair/phase5.c          |  870 +++++++++++++---------------------------------
> > >  2 files changed, 254 insertions(+), 619 deletions(-)
> > > 
> > > 
> > ...
> > > diff --git a/repair/phase5.c b/repair/phase5.c
> > > index 8f5e5f59..e69b042c 100644
> > > --- a/repair/phase5.c
> > > +++ b/repair/phase5.c
> > ...
> > > @@ -837,268 +567,202 @@ btnum_to_ops(
...
> > > +
> > > +		/* Ok, now how many free space records do we have? */
> > > +		*nr_extents = count_bno_extents_blocks(agno, &num_freeblocks);
> > > +	} while (1);
> > > +
> > > +	*extra_blocks = (bno_blocks - btr_bno->bload.nr_blocks) +
> > > +			(cnt_blocks - btr_cnt->bload.nr_blocks);
> > > +}
> > > +
> > > +static void
> > > +get_freesp_data(
> > > +	struct xfs_btree_cur		*cur,
> > > +	struct extent_tree_node		*bno_rec,
> > > +	xfs_agblock_t			*freeblks)
> > > +{
> > > +	struct xfs_alloc_rec_incore	*arec = &cur->bc_rec.a;
> > > +
> > > +	arec->ar_startblock = bno_rec->ex_startblock;
> > > +	arec->ar_blockcount = bno_rec->ex_blockcount;
> > > +	if (freeblks)
> > > +		*freeblks += bno_rec->ex_blockcount;
> > > +}
> > > +
> > > +/* Grab one bnobt record. */
> > > +static int
> > > +get_bnobt_record(
> > > +	struct xfs_btree_cur		*cur,
> > > +	void				*priv)
> > > +{
> > > +	struct bt_rebuild		*btr = priv;
> > > +
> > > +	get_freesp_data(cur, btr->bno_rec, btr->freeblks);
> > > +	btr->bno_rec = findnext_bno_extent(btr->bno_rec);
> > 
> > We should probably check for NULL here even if we don't expect it to
> > happen.
> > 
> > Also, this logic where we load the current pointer and get the next had
> > my eyes crossed for a bit. Can we just check the pointer state here to
> > determine whether to find the first record or the next?
> 
> Hm, I did that, and the logic ends up more cluttered looking:
> 
> 	if (cur->bc_btnum == XFS_BTNUM_BNO) {
> 		if (btr->bno_rec == NULL)
> 			btr->bno_rec = findfirst_bno_extent(agno);
> 		else
> 			btr->bno_rec = findnext_bno_extent(btr->bno_rec);
> 	} else {
> 		if (btr->bno_rec == NULL)
> 			btr->bno_rec = findfirst_bcnt_extent(agno);
> 		else
> 			btr->bno_rec = findnext_bcnt_extent(cur->bc_ag.agno,
> 							    btr->bno_rec);
> 	}
> 

Yeah, I figured this approach would boil down to something like this in
the record feeding code. For one, I'd stick that in a function with a
brief comment. Otherwise, despite the cluttered look, the logic itself
is simple and much easier to read IMO. I had to stare at the current
code for a few minutes before I recognized what was going on, then read
through the iteration thing again to make sure I had it correct. This
eliminates that confusion entirely and thus simplifies the broader
implementation. I think it also facilitates reducing the amount of code
even more if we can generalize the build_*() functions a bit more..

> OTOH I think you do have a good point that splitting the code that sets
> bno_rec across three functions is not easy to understand.
> 

Exactly.

> > Finally, I notice we duplicate the build_[bno|cnt]bt() functions for
> > what amounts to a different get_record() callback specific to the tree
> > type. If we also generalize get_record() to use the tree type (it
> > already has the cursor), it seems that much of the duplication can be
> > reduced and the logic simplified.
> 
> Ok, done.
> 
> > > +	return 0;
> > > +}
> > > +
> > > +/* Rebuild a free space by block number btree. */
> > > +static void
> > > +build_bnobt(
> > > +	struct repair_ctx	*sc,
> > > +	xfs_agnumber_t		agno,
> > > +	struct bt_rebuild	*btr_bno,
> > > +	xfs_agblock_t		*freeblks)
> > > +{
> > > +	int			error;
> > > +
> > > +	*freeblks = 0;
> > > +	btr_bno->bload.get_record = get_bnobt_record;
> > > +	btr_bno->bload.claim_block = rebuild_claim_block;
> > 
> > Note that one of these callbacks is introduced in this patch and the
> > other in the previous. I've no issue with splitting the patches as noted
> > in the previous patch, but these should probably be consistent one way
> > or the other.
> 
> All the btree types set ->claim_block to rebuild_claim_block, whereas
> ->get_record is specific to the particular type of btree.
> 

Ok, seems more reasonable in hindsight.

Brian

> > > +	btr_bno->bno_rec = findfirst_bno_extent(agno);
> > > +	btr_bno->freeblks = freeblks;
> > > +
> > > +	error = -libxfs_trans_alloc_empty(sc->mp, &sc->tp);
> > > +	if (error)
> > > +		do_error(
> > > +_("Insufficient memory to construct bnobt rebuild transaction.\n"));
> > > +
> > > +	/* Add all observed bnobt records. */
> > > +	error = -libxfs_btree_bload(btr_bno->cur, &btr_bno->bload, btr_bno);
> > > +	if (error)
> > > +		do_error(
> > > +_("Error %d while creating bnobt btree for AG %u.\n"), error, agno);
> > > +
> > > +	/* Since we're not writing the AGF yet, no need to commit the cursor */
> > > +	libxfs_btree_del_cursor(btr_bno->cur, 0);
> > > +	error = -libxfs_trans_commit(sc->tp);
> > > +	if (error)
> > > +		do_error(
> > > +_("Error %d while writing bnobt btree for AG %u.\n"), error, agno);
> > > +	sc->tp = NULL;
> > > +}
> > > +
> > ...
> > > @@ -2354,48 +2041,14 @@ build_agf_agfl(
> > >  			freelist[i] = cpu_to_be32(NULLAGBLOCK);
> > >  	}
> > >  
> > > -	/*
> > > -	 * do we have left-over blocks in the btree cursors that should
> > > -	 * be used to fill the AGFL?
> > > -	 */
> > > -	if (bno_bt->num_free_blocks > 0 || bcnt_bt->num_free_blocks > 0)  {
> > > -		/*
> > > -		 * yes, now grab as many blocks as we can
> > > -		 */
> > > -		i = 0;
> > > -		while (bno_bt->num_free_blocks > 0 && i < libxfs_agfl_size(mp))
> > > -		{
> > > -			freelist[i] = cpu_to_be32(
> > > -					get_next_blockaddr(agno, 0, bno_bt));
> > > -			i++;
> > > -		}
> > > -
> > > -		while (bcnt_bt->num_free_blocks > 0 && i < libxfs_agfl_size(mp))
> > > -		{
> > > -			freelist[i] = cpu_to_be32(
> > > -					get_next_blockaddr(agno, 0, bcnt_bt));
> > > -			i++;
> > > -		}
> > > -		/*
> > > -		 * now throw the rest of the blocks away and complain
> > > -		 */
> > > -		while (bno_bt->num_free_blocks > 0) {
> > > -			fsb = XFS_AGB_TO_FSB(mp, agno,
> > > -					get_next_blockaddr(agno, 0, bno_bt));
> > > -			error = slab_add(lost_fsb, &fsb);
> > > -			if (error)
> > > -				do_error(
> > > -_("Insufficient memory saving lost blocks.\n"));
> > > -		}
> > > -		while (bcnt_bt->num_free_blocks > 0) {
> > > -			fsb = XFS_AGB_TO_FSB(mp, agno,
> > > -					get_next_blockaddr(agno, 0, bcnt_bt));
> > > -			error = slab_add(lost_fsb, &fsb);
> > > -			if (error)
> > > -				do_error(
> > > -_("Insufficient memory saving lost blocks.\n"));
> > > -		}
> > > +	/* Fill the AGFL with leftover blocks or save them for later. */
> > > +	i = 0;
> > 
> > How about agfl_idx or something since this variable is passed down into
> > a function?
> 
> The 'i' variable existed before this patch.  I can add a new cleanup to
> fix the name...
> 
> > > +	freelist = xfs_buf_to_agfl_bno(agfl_buf);
> > > +	fill_agfl(btr_bno, freelist, &i);
> > > +	fill_agfl(btr_cnt, freelist, &i);
> > >  
> > > +	/* Set the AGF counters for the AGFL. */
> > > +	if (i > 0) {
> > >  		agf->agf_flfirst = 0;
> > >  		agf->agf_fllast = cpu_to_be32(i - 1);
> > >  		agf->agf_flcount = cpu_to_be32(i);
> > ...
> > > @@ -2650,9 +2283,9 @@ _("unable to rebuild AG %u.  Not enough free space in on-disk AG.\n"),
> > >  	/*
> > >  	 * set up agf and agfl
> > >  	 */
> > > -	build_agf_agfl(mp, agno, &bno_btree_curs,
> > > -		&bcnt_btree_curs, freeblks1, extra_blocks,
> > > -		&rmap_btree_curs, &refcnt_btree_curs, lost_fsb);
> > > +	build_agf_agfl(mp, agno, &btr_bno, &btr_cnt, freeblks1, extra_blocks,
> > > +			&rmap_btree_curs, &refcnt_btree_curs, lost_fsb);
> > > +
> > 
> > I was trying to figure out what extra_blocks was used for and noticed
> > that it's not even used by this function. Perhaps a preceding cleanup
> > patch to drop it?
> 
> Ok.
> 
> > >  	/*
> > >  	 * build inode allocation tree.
> > >  	 */
> > > @@ -2674,15 +2307,14 @@ _("unable to rebuild AG %u.  Not enough free space in on-disk AG.\n"),
> > >  	/*
> > >  	 * tear down cursors
> > >  	 */
> > > -	finish_cursor(&bno_btree_curs);
> > > -	finish_cursor(&ino_btree_curs);
> > 
> > Looks like this one shouldn't be going away here.
> 
> Oops, good catch.
> 
> --D
> 
> > Brian
> > 
> > > +	finish_rebuild(mp, &btr_bno, lost_fsb);
> > > +	finish_rebuild(mp, &btr_cnt, lost_fsb);
> > >  	if (xfs_sb_version_hasrmapbt(&mp->m_sb))
> > >  		finish_cursor(&rmap_btree_curs);
> > >  	if (xfs_sb_version_hasreflink(&mp->m_sb))
> > >  		finish_cursor(&refcnt_btree_curs);
> > >  	if (xfs_sb_version_hasfinobt(&mp->m_sb))
> > >  		finish_cursor(&fino_btree_curs);
> > > -	finish_cursor(&bcnt_btree_curs);
> > >  
> > >  	/*
> > >  	 * release the incore per-AG bno/bcnt trees so the extent nodes
> > > 
> > 
>