Bug#55462 Performance schema: reduce the overhead of PFS_events_waits::m_wait_class

This is a performance improvement fix.

Removed the "volatile" property of PFS_events_waits::m_wait_class.
Simplified the code accordingly.

storage/perfschema/pfs.cc

@@ -1117,7 +1117,6 @@ get_thread_mutex_locker_v1(PSI_mutex_locker_state *state,
   }
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_mutex= pfs_mutex;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1163,7 +1162,6 @@ get_thread_rwlock_locker_v1(PSI_rwlock_locker_state *state,
   }
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_rwlock= pfs_rwlock;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1222,7 +1220,6 @@ get_thread_cond_locker_v1(PSI_cond_locker_state *state,
   }
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_cond= pfs_cond;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1267,7 +1264,6 @@ get_thread_table_locker_v1(PSI_table_locker_state *state,
   }
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_table= pfs_table;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1320,7 +1316,6 @@ get_thread_file_name_locker_v1(PSI_file_locker_state *state,
 
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_file= pfs_file;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1372,7 +1367,6 @@ get_thread_file_stream_locker_v1(PSI_file_locker_state *state,
   }
   PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
     [pfs_thread->m_wait_locker_count];
-  pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
   pfs_locker->m_target.m_file= pfs_file;
   pfs_locker->m_waits_current.m_thread= pfs_thread;
@@ -1441,7 +1435,6 @@ get_thread_file_descriptor_locker_v1(PSI_file_locker_state *state,
       }
       PFS_wait_locker *pfs_locker= &pfs_thread->m_wait_locker_stack
         [pfs_thread->m_wait_locker_count];
-      pfs_locker->m_waits_current.m_wait_class= NO_WAIT_CLASS;
 
       pfs_locker->m_target.m_file= pfs_file;
       pfs_locker->m_waits_current.m_thread= pfs_thread;

storage/perfschema/pfs_events_waits.cc

@@ -80,25 +80,10 @@ void cleanup_events_waits_history_long(void)
   events_waits_history_long_array= NULL;
 }
 
-static void copy_events_waits(PFS_events_waits *dest,
-                              const PFS_events_waits *source)
+static inline void copy_events_waits(PFS_events_waits *dest,
+                                     const PFS_events_waits *source)
 {
-  /* m_wait_class must be the first member of PFS_events_waits. */
-  compile_time_assert(offsetof(PFS_events_waits, m_wait_class) == 0);
-
-  char* dest_body= (reinterpret_cast<char*> (dest)) + sizeof(events_waits_class);
-  const char* source_body= (reinterpret_cast<const char*> (source))
-    + sizeof(events_waits_class);
-
-  /* See comments in table_events_waits_common::make_row(). */
-
-  /* Signal readers they are about to read garbage ... */
-  dest->m_wait_class= NO_WAIT_CLASS;
-  /* ... that this can generate. */
-  memcpy(dest_body, source_body,
-         sizeof(PFS_events_waits) - sizeof(events_waits_class));
-  /* Signal readers the record is now clean again. */
-  dest->m_wait_class= source->m_wait_class;
+  memcpy(dest, source, sizeof(PFS_events_waits));
 }
 
 /**
@@ -116,9 +101,7 @@ void insert_events_waits_history(PFS_thread *thread, PFS_events_waits *wait)
     causing a potential race condition.
     We are not testing for this and insert a possibly empty record,
     to make this thread (the writer) faster.
-    This is ok, the truncated data will have
-    wait->m_wait_class == NO_WAIT_CLASS,
-    which readers of m_waits_history will filter out.
+    This is ok, the readers of m_waits_history will filter this out.
   */
   copy_events_waits(&thread->m_waits_history[index], wait);
 

storage/perfschema/pfs_events_waits.h

@@ -97,7 +97,7 @@ struct PFS_events_waits
     - TRUNCATE EVENTS_WAITS_HISTORY
     - TRUNCATE EVENTS_WAITS_HISTORY_LONG
   */
-  volatile events_waits_class m_wait_class;
+  events_waits_class m_wait_class;
   /** Executing thread. */
   PFS_thread *m_thread;
   /** Instrument metadata. */

storage/perfschema/table_events_waits.cc

@@ -217,16 +217,8 @@ void table_events_waits_common::make_row(bool thread_own_wait,
     or 8 atomics per recorded event.
     The problem is that we record a *lot* of events ...
 
-    Instead, a *dirty* marking is done using m_wait_class.
-    Using m_wait_class alone does not guarantee anything, it just filters
-    out most of the bad data.
-    This is acceptable because this code is garbage-proof,
-    and won't crash on bad data, only display it,
-    very rarely (which is accepted).
-
-    If a bad record is displayed, it's a very transient failure:
-    the next select * from EVENTS_WAITS_CURRENT/_HISTORY/_HISTORY_LONG will
-    show clean data again.
+    This code is prepared to accept *dirty* records,
+    and sanitizes all the data before returning a row.
   */
 
   m_row.m_thread_internal_id= safe_thread->m_thread_internal_id;