"buf_flush_LRU is lazy" | buf_flush_common() calls buf_dblwr_flush_buffered_writes() even when no pages flushed

[Laurynas Biveinis]

This is a copy of http://bugs.mysql.com/bug.php?id=69170, namely the reported issue

"1) buf_flush_common always calls buf_dblwr_flush_buffered_writes. Shouldn't this only be done when page_count > 0?"

which later the bug was kept open about, and

[27 Sep 10:15] Laurynas Biveinis

Re. Mark's problem 1), merely skipping buf_dblwr_flush_buffered_writes() if page_count == 0 will be only a partial fix in the presence of compressed pages for LRU flushes, because this value will originate from buf_do_LRU_batch():

 if (buf_LRU_evict_from_unzip_LRU(buf_pool)) {
  count += buf_free_from_unzip_LRU_list_batch(buf_pool, max);
 }

 if (max > count) {
  count += buf_flush_LRU_list_batch(buf_pool, max - count);
 }

Notice a mixup of evicted and flushed pages. In the case we will satisfy this LRU flush batch purely by unzip_LRU eviction and no flushing, we will still get page_count > 0 in buf_flush_common().

We are exploring a fix here and elsewhere of replacing single value counters with two-field structs for flushed and evicted pages.

[Laurynas Biveinis (laurynas-biveinis)]

See also http://bugs.mysql.com/bug.php?id=71411

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-2430