MDEV-34705: Binlog in Engine: Fix losing last part of event group > trx cache in-memory buffer

Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>

storage/innobase/fsp/fsp0fsp.cc

@@ -4150,6 +4150,22 @@ void fsp_binlog_write_cache(IO_CACHE *cache, size_t main_size, mtr_t *mtr)
   */
   size_t remain= my_b_tell(cache);
   ut_ad(remain > main_size);
+  if (cache->pos_in_file > 0) {
+    /*
+      ToDo: A limitation in mysys IO_CACHE. If I change (reinit_io_cache())
+      the cache from WRITE_CACHE to READ_CACHE without seeking out of the
+      current buffer, then the cache will not be flushed to disk (which is
+      good for small cache that fits completely in buffer). But then if I
+      later my_b_seek() or reinit_io_cache() it again and seek out of the
+      current buffer, the buffered data will not be flushed to the file
+      because the cache is now a READ_CACHE! The result is that the end of the
+      cache will be lost if the cache doesn't fit in memory.
+
+      So for now, have to do this somewhat in-elegant conditional flush
+      myself.
+    */
+    flush_io_cache(cache);
+  }
   /* Start with the GTID event, which is put at the end of the IO_CACHE. */
   my_bool res= reinit_io_cache(cache, READ_CACHE, main_size, 0, 0);
   ut_a(!res /* ToDo: Error handling. */);