Merge gbichot@bk-internal.mysql.com:/home/bk/mysql-maria

into  gbichot4.local:/home/mysql_src/mysql-maria-for-undo-phase


storage/maria/ma_recovery.c:
  Auto merged

mysql-test/r/maria.result

@@ -1976,7 +1976,7 @@ drop table t1;
 show variables like 'maria%';
 Variable_name	Value
 maria_block_size	8192
-maria_checkpoint_frequency	30
+maria_checkpoint_frequency	0
 maria_max_sort_file_size	9223372036853727232
 maria_pagecache_age_threshold	300
 maria_pagecache_buffer_size	8384512

storage/maria/ha_maria.cc

@@ -95,7 +95,8 @@ static MYSQL_SYSVAR_ULONG(checkpoint_frequency, maria_checkpoint_frequency,
        PLUGIN_VAR_RQCMDARG,
        "Frequency of automatic checkpoints, in seconds;"
        " 0 means 'no checkpoints'.",
-       NULL, update_checkpoint_frequency, 30, 0, UINT_MAX, 1);
+                          /* disabled for now */
+       NULL, update_checkpoint_frequency, 0, 0, UINT_MAX, 1);
 
 static MYSQL_SYSVAR_ULONGLONG(max_sort_file_size,
        maria_max_temp_length, PLUGIN_VAR_RQCMDARG,
@@ -2541,6 +2542,7 @@ static struct st_mysql_sys_var* system_variables[]= {
    checkpoint frequency. So when the user wants to modify it, we stop and
    restart the thread.
 */
+
 static void update_checkpoint_frequency(MYSQL_THD thd,
                                         struct st_mysql_sys_var *var,
                                         void *var_ptr, void *save)

storage/maria/ma_pagecache.c

@@ -841,7 +841,7 @@ static int flush_all_key_blocks(PAGECACHE *pagecache)
         KEYCACHE_DBUG_ASSERT(cnt <= pagecache->blocks_used);
 #endif
         if (flush_pagecache_blocks_int(pagecache, &block->hash_link->file,
-                                       FLUSH_RELEASE))
+                                       FLUSH_RELEASE, NULL, NULL))
           return 1;
         break;
       }
@@ -3489,7 +3489,7 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
   pagecache_pthread_mutex_unlock(&pagecache->cache_lock);
   /*
      As all blocks referred in 'cache' are marked by PCBLOCK_IN_FLUSH
-     we are guarantied no thread will change them
+     we are guaranteed that no thread will change them
   */
   qsort((uchar*) cache, count, sizeof(*cache), (qsort_cmp) cmp_sec_link);
 
@@ -3506,6 +3506,8 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
       DBUG_PRINT("info", ("block: %u (0x%lx)  pinned",
                           PCBLOCK_NUMBER(pagecache, block), (ulong)block));
       PCBLOCK_INFO(block);
+      /* undo the mark put by flush_pagecache_blocks_int(): */
+      block->status&= ~PCBLOCK_IN_FLUSH;
       last_errno= -1;
       unreg_request(pagecache, block, 1);
       continue;
@@ -3573,11 +3575,15 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
 
 
 /**
-   @brief flush all key blocks for a file to disk but don't do any mutex locks
+   @brief flush all blocks for a file to disk but don't do any mutex locks
 
    @param  pagecache       pointer to a pagecache data structure
    @param  file            handler for the file to flush to
    @param  flush_type      type of the flush
+   @param  filter          optional function which tells what blocks to flush;
+                           can be non-NULL only if FLUSH_KEEP or FLUSH_FORCE_WRITE.
+   @param  filter_arg      an argument to pass to 'filter'. Information about
+                           the block will be passed too.
 
    @note
      This function doesn't do any mutex locks because it needs to be called
@@ -3591,7 +3597,9 @@ static int flush_cached_blocks(PAGECACHE *pagecache,
 
 static int flush_pagecache_blocks_int(PAGECACHE *pagecache,
                                       PAGECACHE_FILE *file,
-                                      enum flush_type type)
+                                      enum flush_type type,
+                                      PAGECACHE_FLUSH_FILTER filter,
+                                      void *filter_arg)
 {
   PAGECACHE_BLOCK_LINK *cache_buff[FLUSH_CACHE],**cache;
   int last_errno= 0;
@@ -3622,9 +3630,29 @@ static int flush_pagecache_blocks_int(PAGECACHE *pagecache,
 
     if (type != FLUSH_IGNORE_CHANGED)
     {
-      /*
+      /**
          Count how many key blocks we have to cache to be able
-         to flush all dirty pages with minimum seek moves
+         to flush all dirty pages with minimum seek moves.
+
+         @todo RECOVERY BUG
+         We will soon here put code to wait if another thread is flushing the
+         same file, to avoid concurrency bugs. Examples of concurrency bugs
+         which happened without serialization:
+         - assume maria_chk_size() (via CHECK TABLE) happens
+         concurrently with Checkpoint: Checkpoint may be flushing a page, and
+         maria_chk_size() wants to flush this page too so gets an error
+         because Checkpoint pinned this page. Such error leads to marking the
+         table corrupted.
+         - assume maria_close() happens concurrently with Checkpoint:
+         Checkpoint may be flushing a page, and maria_close() flushes this
+         page too with FLUSH_RELEASE: the FLUSH_RELEASE will cause a
+         free_block() which assumes the page is in the LRU, but it is not (as
+         Checkpoint is flushing it). Crash.
+         - assume two flushes of the same file happen concurrently (like
+         above), and a third thread is pushing a page of this file out of the
+         LRU and runs first. Then one flusher will remove the page from
+         changed_blocks[] and put it in its first_in_switch, so the other
+         flusher will not see the page at all and return too early.
       */
       for (block= pagecache->changed_blocks[FILE_HASH(*file)] ;
            block;
@@ -3659,7 +3687,19 @@ restart:
       KEYCACHE_DBUG_ASSERT(cnt <= pagecache->blocks_used);
 #endif
       next= block->next_changed;
-      if (block->hash_link->file.file == file->file)
+      if (block->hash_link->file.file != file->file)
+        continue;
+      if (filter != NULL)
+      {
+        int filter_res= (*filter)(block->type, block->hash_link->pageno,
+                                  block->rec_lsn, filter_arg);
+        DBUG_PRINT("info",("filter returned %d", filter_res));
+        if (filter_res == FLUSH_FILTER_SKIP_TRY_NEXT)
+          continue;
+        if (filter_res == FLUSH_FILTER_SKIP_ALL)
+          break;
+        DBUG_ASSERT(filter_res == FLUSH_FILTER_OK);
+      }
       {
         /*
            Mark the block with BLOCK_IN_FLUSH in order not to let
@@ -3775,6 +3815,11 @@ restart:
     /* The following happens very seldom */
     if (! (type == FLUSH_KEEP || type == FLUSH_FORCE_WRITE))
     {
+      /*
+        this code would free all blocks while filter maybe handled only a
+        few, that is not possible.
+      */
+      DBUG_ASSERT(filter == NULL);
 #if defined(PAGECACHE_DEBUG)
       cnt=0;
 #endif
@@ -3810,23 +3855,27 @@ restart:
 }
 
 
-/*
-  Flush all blocks for a file to disk
-
-  SYNOPSIS
+/**
+   @brief flush all blocks for a file to disk
 
-    flush_pagecache_blocks()
-      pagecache            pointer to a page cache data structure
-      file                handler for the file to flush to
-      flush_type          type of the flush
+   @param  pagecache       pointer to a pagecache data structure
+   @param  file            handler for the file to flush to
+   @param  flush_type      type of the flush
+   @param  filter          optional function which tells what blocks to flush;
+                           can be non-NULL only if FLUSH_KEEP or FLUSH_FORCE_WRITE.
+   @param  filter_arg      an argument to pass to 'filter'. Information about
+                           the block will be passed too.
 
-  RETURN
-    0  OK
-    1  error
+   @return Operation status
+     @retval 0      OK
+     @retval 1      Error
 */
 
-int flush_pagecache_blocks(PAGECACHE *pagecache,
-                           PAGECACHE_FILE *file, enum flush_type type)
+int flush_pagecache_blocks_with_filter(PAGECACHE *pagecache,
+                                       PAGECACHE_FILE *file,
+                                       enum flush_type type,
+                                       PAGECACHE_FLUSH_FILTER filter,
+                                       void *filter_arg)
 {
   int res;
   DBUG_ENTER("flush_pagecache_blocks");
@@ -3836,7 +3885,7 @@ int flush_pagecache_blocks(PAGECACHE *pagecache,
     DBUG_RETURN(0);
   pagecache_pthread_mutex_lock(&pagecache->cache_lock);
   inc_counter_for_resize_op(pagecache);
-  res= flush_pagecache_blocks_int(pagecache, file, type);
+  res= flush_pagecache_blocks_int(pagecache, file, type, filter, filter_arg);
   dec_counter_for_resize_op(pagecache);
   pagecache_pthread_mutex_unlock(&pagecache->cache_lock);
   DBUG_RETURN(res);

storage/maria/ma_pagecache.h

@@ -161,6 +161,18 @@ typedef struct st_pagecache
   my_bool in_init;		/* Set to 1 in MySQL during init/resize     */
 } PAGECACHE;
 
+/** @brief Return values for PAGECACHE_FLUSH_FILTER */
+enum pagecache_flush_filter_result
+{
+  FLUSH_FILTER_SKIP_TRY_NEXT= 0,/**< skip page and move on to next one */
+  FLUSH_FILTER_OK,              /**< flush page and move on to next one */
+  FLUSH_FILTER_SKIP_ALL         /**< skip page and all next ones */
+};
+/** @brief a filter function type for flush_pagecache_blocks_with_filter() */
+typedef enum pagecache_flush_filter_result
+(*PAGECACHE_FLUSH_FILTER)(enum pagecache_page_type type, pgcache_page_no_t page,
+                          LSN rec_lsn, void *arg);
+
 /* The default key cache */
 extern PAGECACHE dflt_pagecache_var, *dflt_pagecache;
 
@@ -228,9 +240,13 @@ extern void pagecache_unpin(PAGECACHE *pagecache,
 extern void pagecache_unpin_by_link(PAGECACHE *pagecache,
                                     PAGECACHE_BLOCK_LINK *link,
                                     LSN lsn);
-extern int flush_pagecache_blocks(PAGECACHE *keycache,
-                                  PAGECACHE_FILE *file,
-                                  enum flush_type type);
+#define flush_pagecache_blocks(A,B,C)                   \
+  flush_pagecache_blocks_with_filter(A,B,C,NULL,NULL)
+extern int flush_pagecache_blocks_with_filter(PAGECACHE *keycache,
+                                              PAGECACHE_FILE *file,
+                                              enum flush_type type,
+                                              PAGECACHE_FLUSH_FILTER filter,
+                                              void *filter_arg);
 extern my_bool pagecache_delete(PAGECACHE *pagecache,
                                 PAGECACHE_FILE *file,
                                 pgcache_page_no_t pageno,

storage/maria/ma_recovery.c

@@ -2256,6 +2256,13 @@ static int close_all_tables(void)
     next_open= list_element->next;
     info= (MARIA_HA*)list_element->data;
     pthread_mutex_unlock(&THR_LOCK_maria); /* ok, UNDO phase not online yet */
+    /*
+      Tables which we see here are exactly those which were open at time of
+      crash. They might have open_count>0 as Checkpoint maybe flushed their
+      state while they were used. As Recovery corrected them, don't alarm the
+      user, don't ask for a table check:
+    */
+    info->s->state.open_count= 0;
     prepare_table_for_close(info, addr);
     error|= maria_close(info);
     pthread_mutex_lock(&THR_LOCK_maria);

storage/maria/ma_test1.c

@@ -20,6 +20,7 @@
 #include <m_string.h>
 #include "ma_control_file.h"
 #include "ma_loghandler.h"
+#include "ma_checkpoint.h"
 #include "trnman.h"
 
 extern PAGECACHE *maria_log_pagecache;
@@ -29,7 +30,7 @@ extern const char *maria_data_root;
 
 static void usage();
 
-static int rec_pointer_size=0, flags[50], testflag;
+static int rec_pointer_size=0, flags[50], testflag, checkpoint;
 static int key_field=FIELD_SKIP_PRESPACE,extra_field=FIELD_SKIP_ENDSPACE;
 static int key_type=HA_KEYTYPE_NUM;
 static int create_flag=0;
@@ -82,7 +83,7 @@ int main(int argc,char *argv[])
       translog_init(maria_data_root, TRANSLOG_FILE_SIZE,
                     0, 0, maria_log_pagecache,
                     TRANSLOG_DEFAULT_FLAGS) ||
-      (transactional && trnman_init(0)))
+      (transactional && (trnman_init(0) || ma_checkpoint_init(FALSE))))
   {
     fprintf(stderr, "Error in initialization");
     exit(1);
@@ -226,6 +227,9 @@ static int run_test(const char *filename)
   if (maria_commit(file) || maria_begin(file))
     goto err;
 
+  if (checkpoint == 1 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
+
   if (testflag == 1)
     goto end;
 
@@ -246,6 +250,9 @@ static int run_test(const char *filename)
     flags[0]=2;
   }
 
+  if (checkpoint == 2 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
+
   if (testflag == 2)
   {
     printf("Terminating after inserts\n");
@@ -307,6 +314,9 @@ static int run_test(const char *filename)
     maria_scan_end(file);
   }
 
+  if (checkpoint == 3 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
+
   if (testflag == 3)
   {
     printf("Terminating after updates\n");
@@ -370,6 +380,9 @@ static int run_test(const char *filename)
     }
   }
 
+  if (checkpoint == 4 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
+
   if (testflag == 4)
   {
     printf("Terminating after deletes\n");
@@ -672,6 +685,8 @@ static void update_record(uchar *record)
 
 static struct my_option my_long_options[] =
 {
+  {"checkpoint", 'H', "Checkpoint at specified stage", (uchar**) &checkpoint,
+   (uchar**) &checkpoint, 0, GET_INT, REQUIRED_ARG, 0, 0, 0, 0, 0, 0},
   {"checksum", 'c', "Undocumented",
    0, 0, 0, GET_NO_ARG, NO_ARG, 0, 0, 0, 0, 0, 0},
 #ifndef DBUG_OFF

storage/maria/ma_test2.c

@@ -28,7 +28,7 @@
 #include "trnman.h"
 #include <m_ctype.h>
 #include <my_bit.h>
-
+#include "ma_checkpoint.h"
 
 #define STANDARD_LENGTH 37
 #define MARIA_KEYS 6
@@ -51,7 +51,7 @@ static	int verbose=0,testflag=0,
             opt_quick_mode=0, transactional= 0, skip_update= 0,
             die_in_middle_of_transaction= 0;
 static int pack_seg=HA_SPACE_PACK,pack_type=HA_PACK_KEY,remove_count=-1;
-static int create_flag= 0, srand_arg= 0;
+static int create_flag= 0, srand_arg= 0, checkpoint= 0;
 static ulong pagecache_size=IO_SIZE*16;
 static enum data_file_type record_type= DYNAMIC_RECORD;
 
@@ -98,7 +98,7 @@ int main(int argc, char *argv[])
       translog_init(maria_data_root, TRANSLOG_FILE_SIZE,
 		    0, 0, maria_log_pagecache,
 		    TRANSLOG_DEFAULT_FLAGS) ||
-      (transactional && trnman_init(0)))
+      (transactional && (trnman_init(0) || ma_checkpoint_init(FALSE))))
   {
     fprintf(stderr, "Error in initialization");
     exit(1);
@@ -240,6 +240,8 @@ int main(int argc, char *argv[])
   maria_begin(file);
   if (testflag == 1)
     goto end;
+  if (checkpoint == 1 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
   if (!silent)
     printf("- Writing key:s\n");
   if (locking)
@@ -302,6 +304,8 @@ int main(int argc, char *argv[])
   }
   if (testflag == 2)
     goto end;
+  if (checkpoint == 2 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
 
   if (write_cacheing)
   {
@@ -353,6 +357,8 @@ int main(int argc, char *argv[])
   }
   if (testflag == 3)
     goto end;
+  if (checkpoint == 3 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
 
   if (!silent)
     printf("- Update\n");
@@ -414,6 +420,8 @@ int main(int argc, char *argv[])
   }
   if (testflag == 4)
     goto end;
+  if (checkpoint == 4 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
 
   for (i=999, dupp_keys=j=0 ; i>0 ; i--)
   {
@@ -824,6 +832,8 @@ int main(int argc, char *argv[])
 
   if (testflag == 5)
     goto end;
+  if (checkpoint == 5 && ma_checkpoint_execute(CHECKPOINT_MEDIUM, FALSE))
+    goto err;
 
   if (!silent)
     printf("- Removing keys\n");
@@ -1057,6 +1067,9 @@ static void get_options(int argc, char **argv)
       if ((first_key=atoi(++pos)) < 0 || first_key >= MARIA_KEYS)
 	first_key=0;
       break;
+    case 'H':
+      checkpoint= atoi(++pos);
+      break;
     case 'k':
       if ((keys=(uint) atoi(++pos)) < 1 ||
 	   keys > (uint) (MARIA_KEYS-first_key))

storage/maria/ma_test_recovery

@@ -126,16 +126,22 @@ echo "Testing the REDO AND UNDO PHASE"
 # Then we run it again and let it exit at T2. Then we compare
 # and expect identity.
 
+for take_checkpoint in "no" "yes"
+do
 for blobs in "" "-b" # we test table without blobs and then table with blobs
 do
   for test_undo in 1 2 3
   do
   # first iteration tests rollback of insert, second tests rollback of delete
-  set -- "ma_test1 $silent -M -T -c -N $blobs" "--testflag=1" "--testflag=2 --test-undo=" "ma_test1 $silent -M -T -c -N $blobs" "--testflag=3" "--testflag=4 --test-undo=" "ma_test1 $silent -M -T -c -N $blobs" "--testflag=2" "--testflag=3 --test-undo=" "ma_test2 $silent -L -K -W -P -M -T -c $blobs" "-t1" "-t2 -u"
+  set -- "ma_test1 $silent -M -T -c -N $blobs -H1" "--testflag=1" "--testflag=2 --test-undo=" "ma_test1 $silent -M -T -c -N $blobs -H2" "--testflag=3" "--testflag=4 --test-undo=" "ma_test1 $silent -M -T -c -N $blobs -H2 " "--testflag=2" "--testflag=3 --test-undo=" "ma_test2 $silent -L -K -W -P -M -T -c $blobs -H1" "-t1" "-t2 -u"
   # -N (create NULL fields) is needed because --test-undo adds it anyway
   while [ $# != 0 ]
     do
     prog=$1
+    if [ "$take_checkpoint" == "no" ]
+        then
+        prog=`echo $prog | sed 's/ -H[0-9]//'`
+        fi
     commit_run_args=$2
     abort_run_args=$3;
     rm -f maria_log.* maria_log_control
@@ -192,6 +198,7 @@ do
   rm -f $table.* $tmp/$table* $tmp/maria_chk_*.txt $tmp/maria_read_log_$table.txt
 done
 done
+done
 
 ) 2>&1 > $tmp/ma_test_recovery.output