mariadb/storage/maria/ma_checkpoint.h
unknown 77017191de WL#3071 - Maria checkpoint
- serializing calls to flush_pagecache_blocks_int() on the same file
to avoid known concurrency bugs
- having that, we can now enable the background thread, as the
flushes it does are now supposedly safe in concurrent situations.
- new type of flush FLUSH_KEEP_LAZY: when the background checkpoint
thread is flushing a packet of dirty pages between two checkpoints,
it uses this flush type, indeed if a file is already being flushed
by another thread it's smarter to move on to the next file than wait.
- maria_checkpoint_frequency renamed to maria_checkpoint_interval.


include/my_sys.h:
  new type of flushing for the page cache: FLUSH_KEEP_LAZY
mysql-test/r/maria.result:
  result update
mysys/mf_keycache.c:
  indentation. No FLUSH_KEEP_LAZY support in key cache.
storage/maria/ha_maria.cc:
  maria_checkpoint_frequency was somehow a hidden part of the
  Checkpoint API and that was not good. Now we have checkpoint_interval,
  local to ha_maria.cc, which serves as container for the user-visible
  maria_checkpoint_interval global variable; setting it calls
  update_checkpoint_interval which passes the new value to
  ma_checkpoint_init(). There is no hiding anymore.
  By default, enable background thread which does checkpoints
  every 30 seconds, and dirty page flush in between. That thread takes
  a checkpoint when it ends, so no need for maria_hton_panic to take one.
  The | is | and not ||, because maria_panic() must always be called.
  frequency->interval.
storage/maria/ma_checkpoint.c:
  Use FLUSH_KEEP_LAZY for background thread when it flushes packets of
  dirty pages between two checkpoints: it is smarter to move on to
  the next file than wait for it to have been completely flushed, which
  may take long.
  Comments about flush concurrency bugs moved from ma_pagecache.c.
  Removing out-of-date comment.
  frequency->interval.
  create_background_thread -> (interval>0).
  In ma_checkpoint_background(), some variables need to be preserved
  between iterations.
storage/maria/ma_checkpoint.h:
  new prototype
storage/maria/ma_pagecache.c:
  - concurrent calls of flush_pagecache_blocks_int() on the same file
  cause bugs (see @note in that function); we fix them by serializing
  in this situation. For that we use a global hash of (file, wqueue).
  When flush_pagecache_blocks_int() starts it looks into the hash,
  using the file as key. If not found, it inserts (file,wqueue) into the
  hash, flushes the file, and finally removes itself from the hash and
  wakes up any waiter in the queue. If found, it adds itself to the
  wqueue and waits.
  - As a by-product, we can remove changed_blocks_is_incomplete
  and replace it by scanning the hash, replace the sleep() by a queue wait.
  - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's
  already being flushed by another thread (even partially), return
  immediately.
storage/maria/ma_pagecache.h:
  In pagecache, a hash of files currently being flushed (i.e. there
  is a call to flush_pagecache_blocks_int() for them).
storage/maria/ma_recovery.c:
  new prototype
storage/maria/ma_test1.c:
  new prototype
storage/maria/ma_test2.c:
  new prototype
2007-10-19 14:15:13 +02:00

81 lines
2.9 KiB
C

/* Copyright (C) 2006,2007 MySQL AB
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA */
/*
WL#3071 Maria checkpoint
First version written by Guilhem Bichot on 2006-04-27.
Does not compile yet.
*/
/* This is the interface of this module. */
typedef enum enum_ma_checkpoint_level {
CHECKPOINT_NONE= 0,
/* just write dirty_pages, transactions table and sync files */
CHECKPOINT_INDIRECT,
/* also flush all dirty pages which were already dirty at prev checkpoint */
CHECKPOINT_MEDIUM,
/* also flush all dirty pages */
CHECKPOINT_FULL
} CHECKPOINT_LEVEL;
C_MODE_START
int ma_checkpoint_init(ulong interval);
void ma_checkpoint_end(void);
int ma_checkpoint_execute(CHECKPOINT_LEVEL level, my_bool no_wait);
C_MODE_END
/**
@brief reads some LSNs with special trickery
If a 64-bit variable transitions between both halves being zero to both
halves being non-zero, and back, this function can be used to do a read of
it (without mutex, without atomic load) which always produces a correct
(though maybe slightly old) value (even on 32-bit CPUs). The value is at
least as new as the latest mutex unlock done by the calling thread.
The assumption is that the system sets both 4-byte halves either at the
same time, or one after the other (in any order), but NOT some bytes of the
first half then some bytes of the second half then the rest of bytes of the
first half. With this assumption, the function can detect when it is
seeing an inconsistent value.
@param LSN pointer to the LSN variable to read
@return LSN part (most significant byte always 0)
*/
#if ( SIZEOF_CHARP >= 8 )
/* 64-bit CPU, 64-bit reads are atomic */
#define lsn_read_non_atomic LSN_WITH_FLAGS_TO_LSN
#else
static inline LSN lsn_read_non_atomic_32(const volatile LSN *x)
{
/*
32-bit CPU, 64-bit reads may give a mixed of old half and new half (old
low bits and new high bits, or the contrary).
*/
for (;;) /* loop until no atomicity problems */
{
/*
Remove most significant byte in case this is a LSN_WITH_FLAGS object.
Those flags in TRN::first_undo_lsn break the condition on transitions so
they must be removed below.
*/
LSN y= LSN_WITH_FLAGS_TO_LSN(*x);
if (likely((y == LSN_IMPOSSIBLE) || LSN_VALID(y)))
return y;
}
}
#define lsn_read_non_atomic(x) lsn_read_non_atomic_32(&x)
#endif