WL#3072 Maria Recovery. Making DDLs durable in Maria:

Sync table files after CREATE (of non-temp table), DROP, RENAME,
TRUNCATE, sync directories and symlinks (for the 3 first commands).
Comments for future log records.
In ma_rename(), if rename of index works and then rename of data fails,
try to undo the rename of the index to leave a consistent state.


mysys/my_symlink.c:
  sync directory after creation of a symbolic link in it, if asked
mysys/my_sync.c:
  comment. Fix for when the file's name has no directory in it.
storage/maria/ma_create.c:
  sync files and links and dirs when creating a non-temporary table.
  Optimizations of the above to reduce syncs in the common cases:
  * if index file and data file have the exact same paths (regular
  and link), sync the directories (of regular and link) only once
  after creating the last file (the data file).
  * don't sync the data file if we didn't write to it (always true
  in our builds).
storage/maria/ma_delete_all.c:
  sync files after truncating a table
storage/maria/ma_delete_table.c:
  sync files and symbolic links and dirs after dropping a table
storage/maria/ma_extra.c:
  a function which wraps the sync of the index file and the sync of the
  data file.
storage/maria/ma_locking.c:
  using a wrapper function
storage/maria/ma_rename.c:
  sync files and symbolic links and dirs after renaming a table.
  If rename of index works and then rename of data fails, try to undo
  the rename of the index to leave a consistent state. That is just a
  try, it may fail...
storage/maria/ma_test3.c:
  warning to not pay attention to this test.
storage/maria/maria_def.h:
  declaration for the function added to ma_extra.c

mysys/my_symlink.c

@@ -85,6 +85,8 @@ int my_symlink(const char *content, const char *linkname, myf MyFlags)
     if (MyFlags & MY_WME)
       my_error(EE_CANT_SYMLINK, MYF(0), linkname, content, errno);
   }
+  else if ((MyFlags & MY_SYNC_DIR) && my_sync_dir_by_file(linkname, MyFlags))
+    result= -1;
   DBUG_RETURN(result);
 #endif /* HAVE_READLINK */
 }

mysys/my_sync.c

@@ -52,7 +52,7 @@ int my_sync(File fd, myf my_flags)
 #if defined(F_FULLFSYNC)
     /*
       In Mac OS X >= 10.3 this call is safer than fsync() (it forces the
-      disk's cache).
+      disk's cache and guarantees ordered writes).
     */
     if (!(res= fcntl(fd, F_FULLFSYNC, 0)))
       break; /* ok */
@@ -89,6 +89,7 @@ int my_sync(File fd, myf my_flags)
 } /* my_sync */
 
 
+static const char cur_dir_name[]= {FN_CURLIB, 0};
 /*
   Force directory information to disk.
 
@@ -107,11 +108,14 @@ int my_sync_dir(const char *dir_name, myf my_flags)
   DBUG_PRINT("my",("Dir: '%s'  my_flags: %d", dir_name, my_flags));
   File dir_fd;
   int res= 0;
+  const char *correct_dir_name;
+  /* Sometimes the path does not contain an explicit directory */
+  correct_dir_name= (dir_name[0] == 0) ? cur_dir_name : dir_name;
   /*
     Syncing a dir may give EINVAL on tmpfs on Linux, which is ok.
     EIO on the other hand is very important. Hence MY_IGNORE_BADFD.
   */
-  if ((dir_fd= my_open(dir_name, O_RDONLY, MYF(my_flags))) >= 0)
+  if ((dir_fd= my_open(correct_dir_name, O_RDONLY, MYF(my_flags))) >= 0)
   {
     if (my_sync(dir_fd, MYF(my_flags | MY_IGNORE_BADFD)))
       res= 2;

storage/maria/ma_create.c

@@ -60,6 +60,8 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
   ulong *rec_per_key_part;
   my_off_t key_root[HA_MAX_POSSIBLE_KEY],key_del[MARIA_MAX_KEY_BLOCK_SIZE];
   MARIA_CREATE_INFO tmp_create_info;
+  my_bool tmp_table= FALSE; /* cache for presence of HA_OPTION_TMP_TABLE */
+  myf     sync_dir=  MY_SYNC_DIR;
   DBUG_ENTER("maria_create");
   DBUG_PRINT("enter", ("keys: %u  columns: %u  uniques: %u  flags: %u",
                       keys, columns, uniques, flags));
@@ -560,7 +562,11 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
 
   /* max_data_file_length and max_key_file_length are recalculated on open */
   if (options & HA_OPTION_TMP_TABLE)
+  {
+    tmp_table= TRUE;
+    sync_dir= 0;
     share.base.max_data_file_length=(my_off_t) ci->data_file_length;
+  }
 
   share.base.min_block_length=
     (share.base.pack_reclength+3 < MARIA_EXTEND_BLOCK_LENGTH &&
@@ -576,7 +582,7 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
   {
     char *iext= strrchr(ci->index_file_name, '.');
     int have_iext= iext && !strcmp(iext, MARIA_NAME_IEXT);
-    if (options & HA_OPTION_TMP_TABLE)
+    if (tmp_table)
     {
       char *path;
       /* chop off the table name, tempory tables use generated name */
@@ -597,8 +603,11 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
     /*
       Don't create the table if the link or file exists to ensure that one
       doesn't accidently destroy another table.
+      Don't sync dir now if the data file has the same path.
     */
-    create_flag=0;
+    create_flag=
+      (ci->data_file_name &&
+       !strcmp(ci->index_file_name, ci->data_file_name)) ? 0 : sync_dir;
   }
   else
   {
@@ -607,8 +616,11 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
                (flags & HA_DONT_TOUCH_DATA) ? MY_RETURN_REAL_PATH : 0) |
                 MY_APPEND_EXT);
     linkname_ptr=0;
-    /* Replace the current file */
-    create_flag=MY_DELETE_OLD;
+    /*
+      Replace the current file.
+      Don't sync dir now if the data file has the same path.
+    */
+    create_flag= MY_DELETE_OLD | (!ci->data_file_name ? 0 : sync_dir);
   }
 
   /*
@@ -627,7 +639,7 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
   }
 
   if ((file= my_create_with_symlink(linkname_ptr, filename, 0, create_mode,
-				    MYF(MY_WME | create_flag))) < 0)
+				    MYF(MY_WME|create_flag))) < 0)
     goto err;
   errpos=1;
 
@@ -653,7 +665,7 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
         char *dext= strrchr(ci->data_file_name, '.');
         int have_dext= dext && !strcmp(dext, MARIA_NAME_DEXT);
 
-        if (options & HA_OPTION_TMP_TABLE)
+        if (tmp_table)
         {
           char *path;
           /* chop off the table name, tempory tables use generated name */
@@ -682,7 +694,7 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
       }
       if ((dfile=
 	   my_create_with_symlink(linkname_ptr, filename, 0, create_mode,
-				  MYF(MY_WME | create_flag))) < 0)
+				  MYF(MY_WME | create_flag | sync_dir))) < 0)
 	goto err;
     }
     errpos=3;
@@ -802,12 +814,18 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
   if (my_chsize(file,(ulong) share.base.keystart,0,MYF(0)))
     goto err;
 
+  if (!tmp_table && my_sync(file, MYF(0)))
+    goto err;
+
   if (! (flags & HA_DONT_TOUCH_DATA))
   {
 #ifdef USE_RELOC
     if (my_chsize(dfile,share.base.min_pack_length*ci->reloc_rows,0,MYF(0)))
       goto err;
+    if (!tmp_table && my_sync(dfile, MYF(0)))
+      goto err;
 #endif
+    /* if !USE_RELOC, there was no write to the file, no need to sync it */
     errpos=2;
     if (my_close(dfile,MYF(0)))
       goto err;
@@ -816,6 +834,19 @@ int maria_create(const char *name,uint keys,MARIA_KEYDEF *keydefs,
   pthread_mutex_unlock(&THR_LOCK_maria);
   if (my_close(file,MYF(0)))
     goto err;
+  /*
+    RECOVERYTODO
+    Write a log record describing the CREATE operation (just the file
+    names, link names, and the full header's content).
+    For this record to be of any use for Recovery, we need the upper
+    MySQL layer to be crash-safe, which it is not now (that would require work
+    using the ddl_log of sql/sql_table.cc); when is is, we should reconsider
+    the moment of writing this log record (before or after op, under
+    THR_LOCK_maria or not...), how to use it in Recovery, and force the log.
+    For now this record is just informative.
+    If operation failed earlier, we clean up in "err:" and the MySQL layer
+    will clean up the frm, so we needn't write anything to the log.
+  */
   my_free((char*) rec_per_key_part,MYF(0));
   DBUG_RETURN(0);
 
@@ -831,14 +862,14 @@ err:
   if (! (flags & HA_DONT_TOUCH_DATA))
     my_delete_with_symlink(fn_format(filename,name,"",MARIA_NAME_DEXT,
                                      MY_UNPACK_FILENAME | MY_APPEND_EXT),
-			   MYF(0));
+			   MYF(sync_dir));
     /* fall through */
   case 1:
     VOID(my_close(file,MYF(0)));
     if (! (flags & HA_DONT_TOUCH_DATA))
       my_delete_with_symlink(fn_format(filename,name,"",MARIA_NAME_IEXT,
                                        MY_UNPACK_FILENAME | MY_APPEND_EXT),
-			     MYF(0));
+			     MYF(sync_dir));
   }
   my_free((char*) rec_per_key_part, MYF(0));
   DBUG_RETURN(my_errno=save_errno);		/* return the fatal errno */

storage/maria/ma_delete_all.c

@@ -30,6 +30,7 @@ int maria_delete_all_rows(MARIA_HA *info)
   {
     DBUG_RETURN(my_errno=EACCES);
   }
+  /* LOCKTODO take X-lock on table here */
   if (_ma_readinfo(info,F_WRLCK,1))
     DBUG_RETURN(my_errno);
   if (_ma_mark_file_changed(info))
@@ -53,9 +54,23 @@ int maria_delete_all_rows(MARIA_HA *info)
     since it was locked then there may be key blocks in the key cache
   */
   flush_key_blocks(share->key_cache, share->kfile, FLUSH_IGNORE_CHANGED);
+  /*
+    RECOVERYTODO Log the two chsize and header modifications and force the
+    log. So that if crash between the two chsize, we finish the work at
+    Recovery. For this scenario:
+    "TRUNCATE TABLE t1; DROP TABLE t1; RENAME TABLE t2 to t1; crash;"
+    Recovery mustn't truncate the new t1, so the log records of TRUNCATE
+    should be applied only if t1 exists and its ZeroDirtyPagesLSN is smaller
+    than the records'. See more comments below.
+  */
   if (my_chsize(info->dfile, 0, 0, MYF(MY_WME)) ||
       my_chsize(share->kfile, share->base.keystart, 0, MYF(MY_WME))  )
     goto err;
+  /*
+    RECOVERYTODO Consider updating ZeroDirtyPagesLSN here. It is
+    not a necessity (it is one only in RENAME commands) but an optional
+    optimization which will allow some REDO skipping at Recovery.
+  */
   VOID(_ma_writeinfo(info,WRITEINFO_UPDATE_KEYFILE));
 #ifdef HAVE_MMAP
   /* Resize mmaped area */
@@ -63,14 +78,25 @@ int maria_delete_all_rows(MARIA_HA *info)
   _ma_remap_file(info, (my_off_t)0);
   rw_unlock(&info->s->mmap_lock);
 #endif
+  /*
+    RECOVERYTODO Until we have the TRUNCATE log record and take it into
+    account for log-low-water-mark calculation and use it in Recovery, we need
+    to sync.
+  */
+  if (_ma_sync_table_files(info))
+    goto err;
   allow_break();			/* Allow SIGHUP & SIGINT */
   DBUG_RETURN(0);
 
 err:
   {
     int save_errno=my_errno;
+    /* RECOVERYTODO log the header modifications */
     VOID(_ma_writeinfo(info,WRITEINFO_UPDATE_KEYFILE));
     info->update|=HA_STATE_WRITTEN;	/* Buffer changed */
+    /* RECOVERYTODO until we log above we have to sync */
+    if (_ma_sync_table_files(info) && !save_errno)
+      save_errno= my_errno;
     allow_break();			/* Allow SIGHUP & SIGINT */
     DBUG_RETURN(my_errno=save_errno);
   }

storage/maria/ma_delete_table.c

@@ -31,6 +31,7 @@ int maria_delete_table(const char *name)
 #ifdef EXTRA_DEBUG
   _ma_check_table_is_closed(name,"delete");
 #endif
+  /* LOCKTODO take X-lock on table here */
 #ifdef USE_RAID
   {
     MARIA_HA *info;
@@ -59,12 +60,22 @@ int maria_delete_table(const char *name)
 #endif /* USE_RAID */
 
   fn_format(from,name,"",MARIA_NAME_IEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
-  if (my_delete_with_symlink(from, MYF(MY_WME)))
+  /*
+    RECOVERYTODO log the two deletes below.
+    Then do the file deletions.
+    For this log record to be of any use for Recovery, we need the upper MySQL
+    layer to be crash-safe in DDLs; when it is we should reconsider the moment
+    of writing this log record, how to use it in Recovery, and force the log.
+    For now this record is only informative.
+  */
+  if (my_delete_with_symlink(from, MYF(MY_WME | MY_SYNC_DIR)))
     DBUG_RETURN(my_errno);
   fn_format(from,name,"",MARIA_NAME_DEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
 #ifdef USE_RAID
   if (raid_type)
-    DBUG_RETURN(my_raid_delete(from, raid_chunks, MYF(MY_WME)) ? my_errno : 0);
+    DBUG_RETURN(my_raid_delete(from, raid_chunks, MYF(MY_WME | MY_SYNC_DIR)) ?
+                my_errno : 0);
 #endif
-  DBUG_RETURN(my_delete_with_symlink(from, MYF(MY_WME)) ? my_errno : 0);
+  DBUG_RETURN(my_delete_with_symlink(from, MYF(MY_WME | MY_SYNC_DIR)) ?
+              my_errno : 0);
 }

storage/maria/ma_extra.c

@@ -316,9 +316,7 @@ int maria_extra(MARIA_HA *info, enum ha_extra_function function, void *extra_arg
     if (share->not_flushed)
     {
       share->not_flushed=0;
-      if (my_sync(share->kfile, MYF(0)))
-	error= my_errno;
-      if (my_sync(info->dfile, MYF(0)))
+      if (_ma_sync_table_files(info))
 	error= my_errno;
       if (error)
       {
@@ -439,3 +437,10 @@ int maria_reset(MARIA_HA *info)
                  HA_STATE_PREV_FOUND);
   DBUG_RETURN(error);
 }
+
+
+int _ma_sync_table_files(const MARIA_HA *info)
+{
+  return (my_sync(info->dfile, MYF(0)) ||
+          my_sync(info->s->kfile, MYF(0)));
+}

storage/maria/ma_locking.c

@@ -103,9 +103,7 @@ int maria_lock_database(MARIA_HA *info, int lock_type)
 	  share->changed=0;
 	  if (maria_flush)
 	  {
-	    if (my_sync(share->kfile, MYF(0)))
-	      error= my_errno;
-	    if (my_sync(info->dfile, MYF(0)))
+            if (_ma_sync_table_files(info))
 	      error= my_errno;
 	  }
 	  else

storage/maria/ma_rename.c

@@ -23,6 +23,7 @@
 int maria_rename(const char *old_name, const char *new_name)
 {
   char from[FN_REFLEN],to[FN_REFLEN];
+  int data_file_rename_error;
 #ifdef USE_RAID
   uint raid_type=0,raid_chunks=0;
 #endif
@@ -32,6 +33,7 @@ int maria_rename(const char *old_name, const char *new_name)
   _ma_check_table_is_closed(old_name,"rename old_table");
   _ma_check_table_is_closed(new_name,"rename new table2");
 #endif
+  /* LOCKTODO take X-lock on table here */
 #ifdef USE_RAID
   {
     MARIA_HA *info;
@@ -48,14 +50,40 @@ int maria_rename(const char *old_name, const char *new_name)
 
   fn_format(from,old_name,"",MARIA_NAME_IEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
   fn_format(to,new_name,"",MARIA_NAME_IEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
-  if (my_rename_with_symlink(from, to, MYF(MY_WME)))
+  /*
+    RECOVERYTODO log the two renames below. Update
+    ZeroDirtyPagesLSN of the table on disk (=> sync the files), this is
+    needed so that Recovery does not pick a wrong table.
+    Then do the file renames.
+    For this log record to be of any use for Recovery, we need the upper MySQL
+    layer to be crash-safe in DDLs; when it is we should reconsider the moment
+    of writing this log record, how to use it in Recovery, and force the log.
+    For now this record is only informative. But ZeroDirtyPagesLSN is
+    critically needed!
+  */
+  if (my_rename_with_symlink(from, to, MYF(MY_WME | MY_SYNC_DIR)))
     DBUG_RETURN(my_errno);
   fn_format(from,old_name,"",MARIA_NAME_DEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
   fn_format(to,new_name,"",MARIA_NAME_DEXT,MY_UNPACK_FILENAME|MY_APPEND_EXT);
 #ifdef USE_RAID
   if (raid_type)
-    DBUG_RETURN(my_raid_rename(from, to, raid_chunks, MYF(MY_WME)) ? my_errno :
-		0);
+    data_file_rename_error= my_raid_rename(from, to, raid_chunks,
+                                           MYF(MY_WME | MY_SYNC_DIR));
+  else
 #endif
-  DBUG_RETURN(my_rename_with_symlink(from, to,MYF(MY_WME)) ? my_errno : 0);
+    data_file_rename_error=
+      my_rename_with_symlink(from, to, MYF(MY_WME | MY_SYNC_DIR));
+  if (data_file_rename_error)
+  {
+    /*
+      now we have a renamed index file and a non-renamed data file, try to
+      undo the rename of the index file.
+    */
+    data_file_rename_error= my_errno;
+    fn_format(from, old_name, "", MARIA_NAME_IEXT, MYF(MY_UNPACK_FILENAME|MY_APPEND_EXT));
+    fn_format(to, new_name, "", MARIA_NAME_IEXT, MYF(MY_UNPACK_FILENAME|MY_APPEND_EXT));
+    my_rename_with_symlink(to, from, MYF(MY_WME | MY_SYNC_DIR));
+  }
+  DBUG_RETURN(data_file_rename_error);
+
 }

storage/maria/ma_test3.c

@@ -65,6 +65,10 @@ int main(int argc,char **argv)
   MY_INIT(argv[0]);
   get_options(argc,argv);
 
+  fprintf(stderr, "WARNING! this program is to test 'external locking'"
+          " (when several processes share a table through file locking)"
+          " which is not supported by Maria at all; expect errors."
+          " We may soon remove this program.\n");
   maria_init();
   bzero((char*) keyinfo,sizeof(keyinfo));
   bzero((char*) recinfo,sizeof(recinfo));

storage/maria/maria_def.h

@@ -741,3 +741,4 @@ int _ma_flush_blocks(HA_CHECK *param, KEY_CACHE *key_cache, File file);
 int _ma_sort_write_record(MARIA_SORT_PARAM *sort_param);
 int _ma_create_index_by_sort(MARIA_SORT_PARAM *info, my_bool no_messages,
                              ulong);
+int _ma_sync_table_files(const MARIA_HA *info);