git: 9917744f7998 - releng/15.2 - Fix direct reads of blocks pending free

From: Colin Percival <cperciva_at_FreeBSD.org>
Date: Sat, 14 Nov 2026 17:26:26 UTC
The branch releng/15.2 has been updated by cperciva:

URL: https://cgit.FreeBSD.org/src/commit/?id=9917744f7998298409f754c98af6591ae297660e

commit 9917744f7998298409f754c98af6591ae297660e
Author:     Wale Zhang <wale.zhang.ftd@gmail.com>
AuthorDate: 2026-11-05 04:31:09 +0000
Commit:     Colin Percival <cperciva@FreeBSD.org>
CommitDate: 2026-11-14 17:25:56 +0000

    Fix direct reads of blocks pending free
    
    Direct reads only check the on-disk block pointer for holes, so they
    can read stale data after a block has been freed but before the free
    has synced. Share buffered read hole detection with direct reads to
    honor pending frees as well.
    
    Keep cached data ahead of hole detection and preserve the transaction
    ordering of block clone and direct write overrides. Retain the block
    pointer recheck after dnode_block_freed() for concurrent syncing.
    
    Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
    Signed-off-by: Wale Zhang <wale.zhang.ftd@gmail.com>
    Closes #19633
    
    Approved by:    re (cperciva)
    Approved by:    kevans (mentor)
    
    (cherry picked from commit 924b8c89c04e324649855b05986f686b58ca4cb9)
    (cherry picked from commit afc7fc2abe807780e6f9cba96ab73105ebd2a07f)
---
 sys/contrib/openzfs/include/sys/dbuf.h      |  1 +
 sys/contrib/openzfs/module/zfs/dbuf.c       | 28 ++++++++++++++++++++--------
 sys/contrib/openzfs/module/zfs/dmu_direct.c |  2 +-
 3 files changed, 22 insertions(+), 9 deletions(-)

diff --git a/sys/contrib/openzfs/include/sys/dbuf.h b/sys/contrib/openzfs/include/sys/dbuf.h
index 0b379ee79caa..a583cbb4d95c 100644
--- a/sys/contrib/openzfs/include/sys/dbuf.h
+++ b/sys/contrib/openzfs/include/sys/dbuf.h
@@ -389,6 +389,7 @@ dbuf_dirty_record_t *dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
 dbuf_dirty_record_t *dbuf_dirty_lightweight(dnode_t *dn, uint64_t blkid,
     dmu_tx_t *tx);
 boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
+boolean_t dbuf_is_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp);
 int dmu_buf_get_bp_from_dbuf(dmu_buf_impl_t *db, blkptr_t **bp);
 int dmu_buf_untransform_direct(dmu_buf_impl_t *db, spa_t *spa);
 void dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
diff --git a/sys/contrib/openzfs/module/zfs/dbuf.c b/sys/contrib/openzfs/module/zfs/dbuf.c
index 9e899e3ea334..1bf22b2639b2 100644
--- a/sys/contrib/openzfs/module/zfs/dbuf.c
+++ b/sys/contrib/openzfs/module/zfs/dbuf.c
@@ -1464,17 +1464,15 @@ dbuf_handle_indirect_hole(void *data, dnode_t *dn, blkptr_t *dbbp)
 }
 
 /*
- * Handle reads on dbufs that are holes, if necessary.  This function
- * requires that the dbuf's mutex is held. Returns success (0) if action
- * was taken, ENOENT if no action was taken.
+ * Check for a hole, including frees which have not yet been synced.
+ * The caller must hold db_mtx to stabilize the dirty records.
  */
-static int
-dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp)
+boolean_t
+dbuf_is_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp)
 {
 	ASSERT(MUTEX_HELD(&db->db_mtx));
-	arc_buf_t *db_data;
 
-	int is_hole = bp == NULL || BP_IS_HOLE(bp);
+	boolean_t is_hole = bp == NULL || BP_IS_HOLE(bp);
 	/*
 	 * For level 0 blocks only, if the above check didn't find a hole,
 	 * consult dnode_block_freed() to check for pending frees.
@@ -1496,7 +1494,21 @@ dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp)
 		}
 	}
 
-	if (is_hole) {
+	return (is_hole);
+}
+
+/*
+ * Handle reads on dbufs that are holes, if necessary.  This function
+ * requires that the dbuf's mutex is held. Returns success (0) if action
+ * was taken, ENOENT if no action was taken.
+ */
+static int
+dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn, blkptr_t *bp)
+{
+	ASSERT(MUTEX_HELD(&db->db_mtx));
+	arc_buf_t *db_data;
+
+	if (dbuf_is_hole(db, dn, bp)) {
 		db_data = dbuf_alloc_arcbuf(db);
 		memset(db_data->b_data, 0, db->db.db_size);
 
diff --git a/sys/contrib/openzfs/module/zfs/dmu_direct.c b/sys/contrib/openzfs/module/zfs/dmu_direct.c
index 5b00698da728..496f4d2afdf2 100644
--- a/sys/contrib/openzfs/module/zfs/dmu_direct.c
+++ b/sys/contrib/openzfs/module/zfs/dmu_direct.c
@@ -298,7 +298,7 @@ dmu_read_abd(dnode_t *dn, uint64_t offset, uint64_t size,
 		 * cached. This will not be considered a direct read for IO
 		 * accounting in the same way that an ARC hit is not counted.
 		 */
-		if (bp == NULL || BP_IS_HOLE(bp) || db->db_state == DB_CACHED) {
+		if (db->db_state == DB_CACHED || dbuf_is_hole(db, dn, bp)) {
 			size_t aoff = offset < db->db.db_offset ?
 			    db->db.db_offset - offset : 0;
 			size_t boff = offset > db->db.db_offset ?