MDEV-26789 log_write_up_to needs mechanism to prevent stalls of async. waiters

Retry write and/or flush, if group_commit_lock::release() indicates that
there are async waiters left, and there is no new group commit lead

Author: vaintroub

storage/innobase/log/log0log.cc

@@ -803,6 +803,9 @@ void log_write_up_to(lsn_t lsn, bool flush_to_disk, bool rotate_key,
     return;
   }
 
+repeat:
+  lsn_t ret_lsn1= 0, ret_lsn2= 0;
+
   if (flush_to_disk &&
       flush_lock.acquire(lsn, callback) != group_commit_lock::ACQUIRED)
     return;
@@ -817,20 +820,32 @@ void log_write_up_to(lsn_t lsn, bool flush_to_disk, bool rotate_key,
     log_write(rotate_key);
 
     ut_a(log_sys.write_lsn == write_lsn);
-    write_lock.release(write_lsn);
+    ret_lsn1= write_lock.release(write_lsn);
   }
 
-  if (!flush_to_disk)
-    return;
-
-  /* Flush the highest written lsn.*/
-  auto flush_lsn = write_lock.value();
-  flush_lock.set_pending(flush_lsn);
-  log_write_flush_to_disk_low(flush_lsn);
-  flush_lock.release(flush_lsn);
+  if (flush_to_disk)
+  {
+    /* Flush the highest written lsn.*/
+    auto flush_lsn = write_lock.value();
+    flush_lock.set_pending(flush_lsn);
+    log_write_flush_to_disk_low(flush_lsn);
+    ret_lsn2= flush_lock.release(flush_lsn);
+
+    log_flush_notify(flush_lsn);
+    DBUG_EXECUTE_IF("crash_after_log_write_upto", DBUG_SUICIDE(););
+  }
 
-  log_flush_notify(flush_lsn);
-  DBUG_EXECUTE_IF("crash_after_log_write_upto", DBUG_SUICIDE(););
+  if (ret_lsn1 || ret_lsn2)
+  {
+    /*
+     There is no new group commit lead, some async waiters could stall.
+     Rerun log_write_up_to(), to prevent that.
+    */
+    lsn= std::max(ret_lsn1, ret_lsn2);
+    static const completion_callback dummy{[](void *) {},nullptr};
+    callback= &dummy;
+    goto repeat;
+  }
 }
 
 /** Write to the log file up to the last log entry.

storage/innobase/log/log0sync.cc

@@ -274,11 +274,11 @@ group_commit_lock::lock_return_code group_commit_lock::acquire(value_type num, c
   return lock_return_code::EXPIRED;
 }
 
-void group_commit_lock::release(value_type num)
+group_commit_lock::value_type group_commit_lock::release(value_type num)
 {
   completion_callback callbacks[1000];
   size_t callback_count = 0;
-
+  value_type ret = 0;
   std::unique_lock<std::mutex> lk(m_mtx);
   m_lock = false;
 
@@ -359,6 +359,12 @@ void group_commit_lock::release(value_type num)
     {
       wakeup_list->m_group_commit_leader=true;
     }
+    else
+    {
+      /* Tell the caller that some pending callbacks left, and he should
+      do something to prevent stalls. This should be a rare situation.*/
+      ret= m_pending_callbacks[0].first;
+    }
   }
 
   lk.unlock();
@@ -371,6 +377,7 @@ void group_commit_lock::release(value_type num)
     next= cur->m_next;
     cur->m_sema.wake();
   }
+  return ret;
 }
 
 #ifndef DBUG_OFF

storage/innobase/log/log0sync.h

@@ -39,16 +39,23 @@ It has a state consisting of
 
 Operations supported on this semaphore
 
-1.acquire(num):
+1.acquire(num, callback):
 - waits until current value exceeds num, or until lock is granted.
+  if running synchronously (callback is nullptr)
 
 - returns EXPIRED if current_value >= num,
-  or ACQUIRED, if current_value < num and lock is granted.
+  or ACQUIRED, if current_value < num and lock is granted,
+  or CALLBACK_QUEUED, if callback was not nullptr, and function
+  would otherwise have to wait
 
 2.release(num)
 - releases lock
 - sets new current value to max(num,current_value)
 - releases some threads waiting in acquire()
+- executes some callbacks
+- might return some lsn, meaning there are some pending
+  callbacks left, and there is no new group commit lead
+  (i.e caller must do something to flush those pending callbacks)
 
 3. value()
 - read current value
@@ -82,7 +89,7 @@ class group_commit_lock
     CALLBACK_QUEUED
   };
   lock_return_code acquire(value_type num, const completion_callback *cb);
-  void release(value_type num);
+  value_type release(value_type num);
   value_type value() const;
   value_type pending() const;
   void set_pending(value_type num);