lmdb: properly sync meta-pages in mdb_sync_env().

Meta-pages may be updated during data-syncing in mdb_sync_env(),
in this case database would be inconsistent.

Check-and-retry if lead txn-id changed during flushing data in mdb_sync_env().

Change-Id: I806e6d3766d028dcd4e6b0591aabc718107056fb
This commit is contained in:
Leo Yuriev 2014-09-19 22:47:19 +04:00
parent fcfe2b7042
commit 84ce8b4cd8

63
mdb.c
View File

@ -2358,17 +2358,22 @@ static int
mdb_env_sync0(MDB_env *env, unsigned int *flags) mdb_env_sync0(MDB_env *env, unsigned int *flags)
{ {
int rc = 0, force; int rc = 0, force;
if (env->me_sync_threshold && env->me_sync_pending >= env->me_sync_threshold) { if (env->me_sync_threshold && env->me_sync_pending >= env->me_sync_threshold)
*flags |= FORCE; *flags |= FORCE;
if (env->me_sync_size != env->me_size)
*flags |= FGREW;
}
force = *flags & FORCE; force = *flags & FORCE;
if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) { if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
if (env->me_sync_size != env->me_size)
*flags |= FGREW;
if (env->me_flags & MDB_WRITEMAP) { if (env->me_flags & MDB_WRITEMAP) {
int mode = ((env->me_flags & MDB_MAPASYNC) && !force) int mode = ((env->me_flags & MDB_MAPASYNC) && !force)
? MS_ASYNC : MS_SYNC; ? MS_ASYNC : MS_SYNC;
if (MDB_MSYNC(env->me_map, env->me_mapsize, mode))
/* LY: skip meta-pages */
size_t data_offset = env->me_os_psize;
while (data_offset < env->me_psize + env->me_psize)
data_offset += env->me_os_psize;
if (MDB_MSYNC(env->me_map + data_offset, env->me_mapsize - data_offset, mode))
rc = ErrCode(); rc = ErrCode();
#ifdef _WIN32 #ifdef _WIN32
else if (mode == MS_SYNC && MDB_FDATASYNC(env->me_fd)) else if (mode == MS_SYNC && MDB_FDATASYNC(env->me_fd))
@ -2397,7 +2402,43 @@ int
mdb_env_sync(MDB_env *env, int force) mdb_env_sync(MDB_env *env, int force)
{ {
unsigned int flags = force ? FORCE | FGREW : 0; unsigned int flags = force ? FORCE | FGREW : 0;
return mdb_env_sync0(env, &flags); MDB_meta *meta;
txnid_t checkpoint;
int rc, lockfree_countdown = 3;
mdb_mutex_t *mutex = NULL;
if (env->me_flags & MDB_RDONLY)
return 0;
do {
if (!mutex && --lockfree_countdown == 0) {
mutex = MDB_MUTEX(env, w);
if (LOCK_MUTEX(rc, env, mutex))
return rc;
}
meta = env->me_metas[ mdb_env_pick_meta(env) ];
checkpoint = meta->mm_txnid;
/* first sync data. */
rc = mdb_env_sync0(env, &flags);
/* then sync meta-pages. */
if (rc == 0 && (env->me_flags & MDB_WRITEMAP)) {
int mode = (!force && (env->me_flags & MDB_MAPASYNC)) ? MS_ASYNC : MS_SYNC;
if (MDB_MSYNC(env->me_map, env->me_psize * 2, mode))
rc = ErrCode();
#ifdef _WIN32
if (rc == 0 && mode == MS_SYNC && MDB_FDATASYNC(env->me_fd))
rc = ErrCode();
#endif
}
} while (rc == 0 && checkpoint != meta->mm_txnid);
if (mutex)
UNLOCK_MUTEX(mutex);
return rc;
} }
/** Back up parent txn's cursors, then grab the originals for tracking */ /** Back up parent txn's cursors, then grab the originals for tracking */
@ -3707,7 +3748,7 @@ mdb_env_write_meta(MDB_txn *txn, int force)
mp->mm_txnid = txn->mt_txnid; mp->mm_txnid = txn->mt_txnid;
if (force || !(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) { if (force || !(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
unsigned meta_size = env->me_psize; unsigned meta_size = env->me_psize;
rc = (!force && (env->me_flags & MDB_MAPASYNC)) ? MS_ASYNC : MS_SYNC; int flags = (!force && (env->me_flags & MDB_MAPASYNC)) ? MS_ASYNC : MS_SYNC;
ptr = env->me_map; ptr = env->me_map;
if (toggle) { if (toggle) {
#ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */ #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
@ -3717,10 +3758,16 @@ mdb_env_write_meta(MDB_txn *txn, int force)
#endif #endif
ptr += meta_size; ptr += meta_size;
} }
if (MDB_MSYNC(ptr, meta_size, rc)) { if (MDB_MSYNC(ptr, meta_size, flags)) {
rc = ErrCode(); rc = ErrCode();
goto fail; goto fail;
} }
#ifdef _WIN32
else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd)) {
rc = ErrCode();
goto fail;
}
#endif
} }
goto done; goto done;
} }