BUG#30947735: NODES ARE CRASHING DURING DISK DATA OBJECT CREATION

Approved-by: Frazer Clement <[email protected]>

We place the extent pages into the list of SL_LOCKED pages before
the data file is ready and before the mutexes are initialised.
This is not good since we crash on trying to lock extent pages.
These pages are not ready for checkpointing, so we can simply
skip them from checkpointing.

Author: Mikael Ronström

storage/ndb/src/kernel/blocks/pgman.cpp

@@ -3256,33 +3256,44 @@ Pgman::process_lcp_locked(Signal* signal, Ptr<Page_entry> ptr)
     {
       Tablespace_client tsman(signal, this, c_tsman, 0, 0, 0, 0);
       jam();
-      tsman.lock_extent_page(ptr.p->m_file_no, ptr.p->m_page_no);
-      if ((ptr.p->m_state & Page_entry::DIRTY) &&
-          !(ptr.p->m_state & Page_entry::PAGEOUT))
+      bool is_file_ready = tsman.is_datafile_ready(ptr.p->m_file_no);
+      if (is_file_ready)
       {
-        jam();
-        Ptr<GlobalPage> org, copy;
-        ndbrequire(m_global_page_pool.seize(copy));
-        m_global_page_pool.getPtr(org, ptr.p->m_real_page_i);
-        memcpy(copy.p, org.p, sizeof(GlobalPage));
-        ptr.p->m_copy_page_i = copy.i;
-
-        ptr.p->m_state |= Page_entry::LCP;
+        /**
+         * An extent page is placed into SL_LOCKED pages before the
+         * data file is ready for use. This means that we haven't even
+         * initialised the mutexes yet and also not initialised all
+         * the extent pages. Avoid checkpointing those pages until
+         * the data file is ready.
+         */
+        tsman.lock_extent_page(ptr.p->m_file_no, ptr.p->m_page_no);
+        if ((ptr.p->m_state & Page_entry::DIRTY) &&
+            !(ptr.p->m_state & Page_entry::PAGEOUT))
+        {
+          jam();
+          Ptr<GlobalPage> org, copy;
+          ndbrequire(m_global_page_pool.seize(copy));
+          m_global_page_pool.getPtr(org, ptr.p->m_real_page_i);
+          memcpy(copy.p, org.p, sizeof(GlobalPage));
+          ptr.p->m_copy_page_i = copy.i;
 
-        DEB_PGMAN_PAGE(("(%u)pageout():extent, page(%u,%u):%u:%x",
-                        instance(),
-                        ptr.p->m_file_no,
-                        ptr.p->m_page_no,
-                        ptr.i,
-                        (unsigned int)ptr.p->m_state));
+          ptr.p->m_state |= Page_entry::LCP;
 
-        pageout(signal, ptr);
-        m_lcp_outstanding++;
-        m_current_lcp_pageouts++;
-        m_available_lcp_pageouts_used++;
-        break_flag = true;
+          DEB_PGMAN_PAGE(("(%u)pageout():extent, page(%u,%u):%u:%x",
+                          instance(),
+                          ptr.p->m_file_no,
+                          ptr.p->m_page_no,
+                          ptr.i,
+                          (unsigned int)ptr.p->m_state));
+
+          pageout(signal, ptr);
+          m_lcp_outstanding++;
+          m_current_lcp_pageouts++;
+          m_available_lcp_pageouts_used++;
+          break_flag = true;
+        }
+        tsman.unlock_extent_page(ptr.p->m_file_no, ptr.p->m_page_no);
       }
-      tsman.unlock_extent_page(ptr.p->m_file_no, ptr.p->m_page_no);
     }
 
     Page_sublist& pl = *m_page_sublist[Page_entry::SL_LOCKED];

storage/ndb/src/kernel/blocks/tsman.cpp

@@ -3428,6 +3428,20 @@ Tsman::client_unlock()
   }
 }
 
+bool Tsman::is_datafile_ready(Uint32 file_no)
+{
+  Ptr<Datafile> file_ptr;
+  Datafile file_key;
+  file_key.m_file_no = file_no;
+  if (m_file_hash.find(file_ptr, file_key))
+  {
+    if (file_ptr.p->m_state == Datafile::FS_CREATING)
+      return false;
+    return true;
+  }
+  return false;
+}
+
 void Tsman::lock_extent_page(Uint32 file_no, Uint32 page_no)
 {
   if (isNdbMtLqh())

storage/ndb/src/kernel/blocks/tsman.hpp

@@ -1,5 +1,5 @@
 /*
-   Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved.
+   Copyright (c) 2005, 2020, Oracle and/or its affiliates. All rights reserved.
 
    This program is free software; you can redistribute it and/or modify
    it under the terms of the GNU General Public License, version 2.0,
@@ -242,6 +242,7 @@ class Tsman : public SimulatedBlock
   void client_unlock();
   void client_lock(Uint32 instance);
   void client_unlock(Uint32 instance);
+  bool is_datafile_ready(Uint32 file_no);
   void lock_extent_page(Uint32 file_no, Uint32 page_no);
   void unlock_extent_page(Uint32 file_no, Uint32 page_no);
   void lock_extent_page(Datafile*, Uint32 page_no);
@@ -413,6 +414,11 @@ class Tablespace_client
    */
   int unmap_page(Local_key*, Uint32 bits);
 
+  /**
+   * Check if datafile is ready for checkpoints.
+   */
+  bool is_datafile_ready(Uint32 file_no);
+
   /**
    * Lock/Unlock extent page to ensure that access to this extent
    * page is serialised.
@@ -553,6 +559,13 @@ Tablespace_client::get_page_free_bits(Local_key *key,
   return m_tsman->get_page_free_bits(m_signal, key, uncommited, commited);
 }
 
+inline
+bool
+Tablespace_client::is_datafile_ready(Uint32 file_no)
+{
+  return m_tsman->is_datafile_ready(file_no);
+}
+
 inline
 void
 Tablespace_client::lock_extent_page(Uint32 file_no, Uint32 page_no)