Hi Brian,

On Mon, Mar 22, 2021 at 07:30:03AM -0400, Brian Foster wrote:

On Fri, Mar 05, 2021 at 10:57:01AM +0800, Gao Xiang wrote:

quoted

This patch introduces a helper to shrink unused space in the last AG
by fixing up the freespace btree.

Also make sure that the per-AG reservation works under the new AG
size. If such per-AG reservation or extent allocation fails, roll
the transaction so the new transaction could cancel without any side
effects.

Signed-off-by: Gao Xiang <redacted>
---

Looks mostly good to me. Some nits..

quoted

 fs/xfs/libxfs/xfs_ag.c | 111 +++++++++++++++++++++++++++++++++++++++++
 fs/xfs/libxfs/xfs_ag.h |   4 +-
 2 files changed, 114 insertions(+), 1 deletion(-)

diff --git a/fs/xfs/libxfs/xfs_ag.c b/fs/xfs/libxfs/xfs_ag.c
index 9331f3516afa..1f6f9e70e1cb 100644
--- a/fs/xfs/libxfs/xfs_ag.c
+++ b/fs/xfs/libxfs/xfs_ag.c

...

quoted

@@ -485,6 +490,112 @@ xfs_ag_init_headers(
 	return error;
 }
 
+int
+xfs_ag_shrink_space(
+	struct xfs_mount	*mp,
+	struct xfs_trans	**tpp,
+	xfs_agnumber_t		agno,
+	xfs_extlen_t		delta)
+{
+	struct xfs_alloc_arg	args = {
+		.tp	= *tpp,
+		.mp	= mp,
+		.type	= XFS_ALLOCTYPE_THIS_BNO,
+		.minlen = delta,
+		.maxlen = delta,
+		.oinfo	= XFS_RMAP_OINFO_SKIP_UPDATE,
+		.resv	= XFS_AG_RESV_NONE,
+		.prod	= 1
+	};
+	struct xfs_buf		*agibp, *agfbp;
+	struct xfs_agi		*agi;
+	struct xfs_agf		*agf;
+	int			error, err2;
+
+	ASSERT(agno == mp->m_sb.sb_agcount - 1);
+	error = xfs_ialloc_read_agi(mp, *tpp, agno, &agibp);
+	if (error)
+		return error;
+
+	agi = agibp->b_addr;
+
+	error = xfs_alloc_read_agf(mp, *tpp, agno, 0, &agfbp);
+	if (error)
+		return error;
+
+	agf = agfbp->b_addr;
+	if (XFS_IS_CORRUPT(mp, agf->agf_length != agi->agi_length))
+		return -EFSCORRUPTED;

Is this check here for a reason? It seems a bit random, so I wonder if
we should just leave the extra verification to buffer verifiers.

It came from Darrick's thought. I'm fine with either way, but I feel
confused if different conflict opinions here:
https://lore.kernel.org/linux-xfs/20210303181931.GB3419940@magnolia/ (local)

quoted

+
+	if (delta >= agi->agi_length)
+		return -EINVAL;
+
+	args.fsbno = XFS_AGB_TO_FSB(mp, agno,
+				    be32_to_cpu(agi->agi_length) - delta);
+
+	/* remove the preallocations before allocation and re-establish then */

The comment is a little confusing. Perhaps something like the following,
if accurate..?

/*
 * Disable perag reservations so it doesn't cause the allocation request
 * to fail. We'll reestablish reservation before we return.
 */

ok, will update in the next version.

quoted

+	error = xfs_ag_resv_free(agibp->b_pag);
+	if (error)
+		return error;
+
+	/* internal log shouldn't also show up in the free space btrees */
+	error = xfs_alloc_vextent(&args);
+	if (!error && args.agbno == NULLAGBLOCK)
+		error = -ENOSPC;
+
+	if (error) {
+		/*
+		 * if extent allocation fails, need to roll the transaction to
+		 * ensure that the AGFL fixup has been committed anyway.
+		 */
+		err2 = xfs_trans_roll(tpp);
+		if (err2)
+			return err2;
+		goto resv_init_out;

So if this fails and the transaction rolls, do we still hold the agi/agf
buffers here? If not, there might be a window of time where it's
possible for some other task to come in and alloc out of the AG without
the perag res being active.

Yeah, I might need to bhold this, will update.

quoted

+	}
+
+	/*
+	 * if successfully deleted from freespace btrees, need to confirm
+	 * per-AG reservation works as expected.
+	 */
+	be32_add_cpu(&agi->agi_length, -delta);
+	be32_add_cpu(&agf->agf_length, -delta);
+
+	err2 = xfs_ag_resv_init(agibp->b_pag, *tpp);
+	if (err2) {
+		be32_add_cpu(&agi->agi_length, delta);
+		be32_add_cpu(&agf->agf_length, delta);
+		if (err2 != -ENOSPC)
+			goto resv_err;
+
+		__xfs_bmap_add_free(*tpp, args.fsbno, delta, NULL, true);
+
+		/*
+		 * Roll the transaction before trying to re-init the per-ag
+		 * reservation. The new transaction is clean so it will cancel
+		 * without any side effects.
+		 */
+		error = xfs_defer_finish(tpp);
+		if (error)
+			return error;
+
+		error = -ENOSPC;
+		goto resv_init_out;
+	}
+	xfs_ialloc_log_agi(*tpp, agibp, XFS_AGI_LENGTH);
+	xfs_alloc_log_agf(*tpp, agfbp, XFS_AGF_LENGTH);
+	return 0;
+
+resv_init_out:
+	err2 = xfs_ag_resv_init(agibp->b_pag, *tpp);
+	if (!err2)
+		return error;
+resv_err:
+	xfs_warn(mp, "Error %d reserving per-AG metadata reserve pool.", err2);
+	xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
+	return err2;
+}
+
 /*
  * Extent the AG indicated by the @id by the length passed in
  */

diff --git a/fs/xfs/libxfs/xfs_ag.h b/fs/xfs/libxfs/xfs_ag.h
index 5166322807e7..41293ebde8da 100644
--- a/fs/xfs/libxfs/xfs_ag.h
+++ b/fs/xfs/libxfs/xfs_ag.h

@@ -24,8 +24,10 @@ struct aghdr_init_data {
 };
 
 int xfs_ag_init_headers(struct xfs_mount *mp, struct aghdr_init_data *id);
+int xfs_ag_shrink_space(struct xfs_mount *mp, struct xfs_trans **tpp,
+			xfs_agnumber_t agno, xfs_extlen_t len);
 int xfs_ag_extend_space(struct xfs_mount *mp, struct xfs_trans *tp,
-			struct aghdr_init_data *id, xfs_extlen_t len);
+			struct aghdr_init_data *id, xfs_extlen_t delta);

This looks misplaced..?

Or maybe this is trying to make the APIs consistent, but the function
definition still uses len as well as the declaration for
_ag_shrink_space() (while the definition of that function uses delta).

FWIW, the name delta tends to suggest a signed value to me based on our
pattern of usage, whereas here it seems like these helpers always want a
positive value (i.e. a length).

Yeah, it's just misplaced, thanks for pointing out, sorry about that.
`delta' name came from, `len' is confusing to Darrick.
https://lore.kernel.org/r/20210303182527.GC3419940@magnolia/ (local)

Thanks,
Gao Xiang

Brian

quoted

 int xfs_ag_get_geometry(struct xfs_mount *mp, xfs_agnumber_t agno,
 			struct xfs_ag_geometry *ageo);
 
-- 
2.27.0