diff --git a/fs/bcachefs/fs-io-buffered.c b/fs/bcachefs/fs-io-buffered.c
index 19d4599918dc..e3a75dcca60c 100644
--- a/fs/bcachefs/fs-io-buffered.c
+++ b/fs/bcachefs/fs-io-buffered.c
@@ -225,11 +225,26 @@ static void bchfs_read(struct btree_trans *trans,
 
 		bch2_read_extent(trans, rbio, iter.pos,
 				 data_btree, k, offset_into_extent, flags);
-		swap(rbio->bio.bi_iter.bi_size, bytes);
+		/*
+		 * Careful there's a landmine here if bch2_read_extent() ever
+		 * starts returning transaction restarts here.
+		 *
+		 * We've changed rbio->bi_iter.bi_size to be "bytes we can read
+		 * from this extent" with the swap call, and we restore it
+		 * below. That restore needs to come before checking for
+		 * errors.
+		 *
+		 * But unlike __bch2_read(), we use the rbio bvec iter, not one
+		 * on the stack, so we can't do the restore right after the
+		 * bch2_read_extent() call: we don't own that iterator anymore
+		 * if BCH_READ_last_fragment is set, since we may have submitted
+		 * that rbio instead of cloning it.
+		 */
 
 		if (flags & BCH_READ_last_fragment)
 			break;
 
+		swap(rbio->bio.bi_iter.bi_size, bytes);
 		bio_advance(&rbio->bio, bytes);
 err:
 		if (ret &&