Change buffer entries for secondary indexes are lost on InnoDB restart

The bug fix MDEV-21069 aimed to avoid a server crash when executing DROP TABLE of a corrupted table. Unfortunately, that fix introduced corruption in the function buf_read_ibuf_merge_pages() and caused wrongful deletion of records from the change buffer for tablespaces whose first page had not been read yet. The following should fix it:

diff --git a/storage/innobase/buf/buf0rea.cc b/storage/innobase/buf/buf0rea.cc

--- a/storage/innobase/buf/buf0rea.cc

+++ b/storage/innobase/buf/buf0rea.cc

@@ -772,13 +772,18 @@ buf_read_ibuf_merge_pages(

 			continue;

 		}

-		if (UNIV_UNLIKELY(page_nos[i] >= space->size)) {

+		ulint size = space->size;

+		if (!size) {

+			size = fil_space_get_size(space->id);

+		}

+

+		if (UNIV_UNLIKELY(page_nos[i] >= size)) {

 			do {

 				ibuf_delete_recs(page_id_t(space_ids[i],

 							   page_nos[i]));

 			} while (++i < n_stored

 				 && space_ids[i - 1] == space_ids[i]

-				 && page_nos[i] >= space->size);

+				 && page_nos[i] >= size);

 			i--;

 next:

 			space->release();

MariaDB Server 10.5 is not affected by this, because this code was removed in MDEV-19514. (However, 10.5 is affected by MDEV-25783 without even involving a server restart.)

Also if crash recovery is needed, those tablespaces for which no redo log was applied could be affected by this bug.

I believe that this kind of corruption can be detected by CHECK TABLE, and it can be fixed by dropping and re-creating the secondary indexes:

CHECK TABLE t QUICK;

ALTER TABLE t DROP INDEX idx1, DROP INDEX idx2;

ALTER TABLE t ADD INDEX idx1(a), ADD INDEX idx2(b);

A work-around would be to ensure that the change buffer is empty during shutdown (set innodb_fast_shutdown=0) or to always run with innodb_change_buffering=none.

This bug was originally reported and analyzed in MDEV-22373, but because MDEV-22373 is partly explained by MDEV-24449 and MDEV-25783, it is clearer to file a separate ticket for this specific case.