/*
* __meta_track_apply --
* Apply the changes in a metadata tracking record.
*/
static int
__meta_track_apply(WT_SESSION_IMPL *session, WT_META_TRACK *trk)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
switch (trk->op) {
case WT_ST_EMPTY: /* Unused slot */
break;
case WT_ST_CHECKPOINT: /* Checkpoint, see above */
btree = trk->dhandle->handle;
bm = btree->bm;
WT_WITH_DHANDLE(session, trk->dhandle,
ret = bm->checkpoint_resolve(bm, session));
break;
case WT_ST_DROP_COMMIT:
if ((ret =
__wt_block_manager_drop(session, trk->a, false)) != 0)
__wt_err(session, ret,
"metadata remove dropped file %s", trk->a);
break;
case WT_ST_LOCK:
WT_WITH_DHANDLE(session, trk->dhandle,
ret = __wt_session_release_btree(session));
break;
case WT_ST_FILEOP:
case WT_ST_REMOVE:
case WT_ST_SET:
break;
}
__meta_track_clear(session, trk);
return (ret);
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, int need_sync, int unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, return now and avoid unnecessary
* metadata checkpoints. For example, this happens if attempting to
* create a data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
return (0);
while (--trk >= trk_orig)
WT_TRET(__meta_track_apply(session, trk, unroll));
/*
* Unroll operations don't need to flush the metadata.
*
* Also, if we don't have the metadata handle (e.g, we're in the
* process of creating the metadata), we can't sync it.
*/
if (unroll || ret != 0 || !need_sync || session->meta_dhandle == NULL)
return (ret);
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED)) {
if (!FLD_ISSET(S2C(session)->txn_logsync,
WT_LOG_DSYNC | WT_LOG_FSYNC))
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_txn_checkpoint_log(session,
0, WT_TXN_LOG_CKPT_SYNC, NULL));
} else {
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_checkpoint(session, NULL));
WT_RET(ret);
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_checkpoint_sync(session, NULL));
}
return (ret);
}
/*
* __compact_worker --
* Function to alternate between checkpoints and compaction calls.
*/
static int
__compact_worker(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
u_int i, loop;
bool didwork;
/*
* Reset the handles' compaction skip flag (we don't bother setting
* or resetting it when we finish compaction, it's simpler to do it
* once, here).
*/
for (i = 0; i < session->op_handle_next; ++i)
session->op_handle[i]->compact_skip = false;
/*
* Perform an initial checkpoint (see this file's leading comment for
* details).
*/
WT_ERR(__compact_checkpoint(session));
/*
* We compact 10% of a file on each pass (but the overall size of the
* file is decreasing each time, so we're not compacting 10% of the
* original file each time). Try 100 times (which is clearly more than
* we need); quit if we make no progress.
*/
for (loop = 0; loop < 100; ++loop) {
/* Step through the list of files being compacted. */
for (didwork = false, i = 0; i < session->op_handle_next; ++i) {
/* Skip objects where there's no more work. */
if (session->op_handle[i]->compact_skip)
continue;
session->compact_state = WT_COMPACT_RUNNING;
WT_WITH_DHANDLE(session,
session->op_handle[i], ret = __wt_compact(session));
WT_ERR(ret);
/* If we did no work, skip this file in the future. */
if (session->compact_state == WT_COMPACT_SUCCESS)
didwork = true;
else
session->op_handle[i]->compact_skip = true;
}
if (!didwork)
break;
/*
* Perform two checkpoints (see this file's leading comment for
* details).
*/
WT_ERR(__compact_checkpoint(session));
WT_ERR(__compact_checkpoint(session));
}
err: session->compact_state = WT_COMPACT_NONE;
return (ret);
}
/*
* __meta_track_unroll --
* Undo the changes in a metadata tracking record.
*/
static int
__meta_track_unroll(WT_SESSION_IMPL *session, WT_META_TRACK *trk)
{
WT_DECL_RET;
switch (trk->op) {
case WT_ST_EMPTY: /* Unused slot */
break;
case WT_ST_CHECKPOINT: /* Checkpoint, see above */
break;
case WT_ST_DROP_COMMIT:
break;
case WT_ST_LOCK: /* Handle lock, see above */
if (trk->created)
F_SET(trk->dhandle, WT_DHANDLE_DISCARD);
WT_WITH_DHANDLE(session, trk->dhandle,
ret = __wt_session_release_btree(session));
break;
case WT_ST_FILEOP: /* File operation */
/*
* For renames, both a and b are set.
* For creates, a is NULL.
* For removes, b is NULL.
*/
if (trk->a != NULL && trk->b != NULL &&
(ret = __wt_fs_rename(session,
trk->b + strlen("file:"), trk->a + strlen("file:"),
true)) != 0)
__wt_err(session, ret,
"metadata unroll rename %s to %s", trk->b, trk->a);
if (trk->a == NULL &&
(ret = __wt_fs_remove(session,
trk->b + strlen("file:"), false)) != 0)
__wt_err(session, ret,
"metadata unroll create %s", trk->b);
/*
* We can't undo removes yet: that would imply
* some kind of temporary rename and remove in
* roll forward.
*/
break;
case WT_ST_REMOVE: /* Remove trk.a */
if ((ret = __wt_metadata_remove(session, trk->a)) != 0)
__wt_err(session, ret,
"metadata unroll remove: %s", trk->a);
break;
case WT_ST_SET: /* Set trk.a to trk.b */
if ((ret = __wt_metadata_update(session, trk->a, trk->b)) != 0)
__wt_err(session, ret,
"metadata unroll update %s to %s", trk->a, trk->b);
break;
}
__meta_track_clear(session, trk);
return (ret);
}
/*
* __wt_schema_release_table --
* Release a table handle.
*/
int
__wt_schema_release_table(WT_SESSION_IMPL *session, WT_TABLE *table)
{
WT_DECL_RET;
WT_WITH_DHANDLE(session, &table->iface,
ret = __wt_session_release_dhandle(session));
return (ret);
}
/*
* __wt_schema_release_table --
* Release a table handle.
*/
int
__wt_schema_release_table(WT_SESSION_IMPL *session, WT_TABLE **tablep)
{
WT_DECL_RET;
WT_TABLE *table;
if ((table = *tablep) == NULL)
return (0);
*tablep = NULL;
WT_WITH_DHANDLE(session, &table->iface,
ret = __wt_session_release_dhandle(session));
return (ret);
}
/*
* __backup_cleanup_handles --
* Release and free all btree handles held by the backup. This is kept
* separate from __backup_stop because it can be called without the
* schema lock held.
*/
static int
__backup_cleanup_handles(WT_SESSION_IMPL *session, WT_CURSOR_BACKUP *cb)
{
WT_CURSOR_BACKUP_ENTRY *p;
WT_DECL_RET;
if (cb->list == NULL)
return (0);
/* Release the handles, free the file names, free the list itself. */
for (p = cb->list; p->name != NULL; ++p) {
if (p->handle != NULL)
WT_WITH_DHANDLE(session, p->handle,
WT_TRET(__wt_session_release_btree(session)));
__wt_free(session, p->name);
}
__wt_free(session, cb->list);
return (ret);
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_SESSION_IMPL *ckpt_session;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, return now and avoid unnecessary
* metadata checkpoints. For example, this happens if attempting to
* create a data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
return (0);
if (unroll) {
while (--trk >= trk_orig)
WT_TRET(__meta_track_unroll(session, trk));
/* Unroll operations don't need to flush the metadata. */
return (ret);
}
/*
* If we don't have the metadata cursor (e.g, we're in the process of
* creating the metadata), we can't sync it.
*/
if (!need_sync || session->meta_cursor == NULL ||
F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto done;
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED)) {
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_txn_checkpoint_log(
session, false, WT_TXN_LOG_CKPT_SYNC, NULL));
WT_RET(ret);
} else {
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
ckpt_session = S2C(session)->meta_ckpt_session;
/*
* If this operation is part of a running transaction, that
* should be included in the checkpoint.
*/
ckpt_session->txn.id = session->txn.id;
F_SET(ckpt_session, WT_SESSION_LOCKED_METADATA);
WT_WITH_METADATA_LOCK(session, ret,
WT_WITH_DHANDLE(ckpt_session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint(ckpt_session, NULL)));
F_CLR(ckpt_session, WT_SESSION_LOCKED_METADATA);
ckpt_session->txn.id = WT_TXN_NONE;
WT_RET(ret);
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint_sync(session, NULL));
WT_RET(ret);
}
done: /* Apply any tracked operations post-commit. */
for (; trk_orig < trk; trk_orig++)
WT_TRET(__meta_track_apply(session, trk_orig));
return (ret);
}
/*
* __sweep --
* Close unused dhandles on the connection dhandle list.
*/
static int
__sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DATA_HANDLE *dhandle, *dhandle_next;
WT_DECL_RET;
time_t now;
conn = S2C(session);
/*
* Session's cache handles unless the session itself is closed, at which
* time the handle reference counts are immediately decremented. Don't
* discard handles that have been open recently.
*/
WT_RET(__wt_seconds(session, &now));
dhandle = SLIST_FIRST(&conn->dhlh);
for (; dhandle != NULL; dhandle = dhandle_next) {
dhandle_next = SLIST_NEXT(dhandle, l);
if (dhandle->session_ref != 0 ||
now - dhandle->timeofdeath <= WT_DHANDLE_SWEEP_WAIT)
continue;
/*
* We have a candidate for closing; if it's open, acquire an
* exclusive lock on the handle and close it (the lock blocks
* threads from opening the handle). We might be blocking an
* open for a fairly long time (over disk I/O), but the handle
* has been quiescent for awhile.
*
* The close can fail if an update cannot be written (updates in
* a no-longer-referenced file might not yet be globally visible
* if sessions have disjoint sets of files open). If the handle
* is busy, skip it, we'll retry the close the next time, after
* the transaction state has progressed.
*/
if (F_ISSET(dhandle, WT_DHANDLE_OPEN)) {
/*
* We don't set WT_DHANDLE_EXCLUSIVE deliberately, we
* want opens to block on us rather than returning an
* EBUSY error to the application.
*/
ret = __wt_try_writelock(session, dhandle->rwlock);
if (ret == EBUSY) {
ret = 0;
continue;
}
WT_RET(ret);
WT_WITH_DHANDLE(session, dhandle,
ret = __wt_conn_btree_sync_and_close(session));
if (ret == EBUSY)
ret = 0;
WT_TRET(__wt_rwunlock(session, dhandle->rwlock));
WT_RET(ret);
}
/*
* Attempt to discard the handle (the called function checks the
* handle-open flag after acquiring appropriate locks, which is
* why we don't do any special handling of EBUSY returns above,
* that path never cleared the handle-open flag.
*/
ret = __wt_conn_dhandle_discard_single(session, dhandle, 0);
if (ret == EBUSY)
ret = 0;
WT_RET(ret);
}
return (0);
}
/*
* __meta_track_apply --
* Apply the changes in a metadata tracking record.
*/
static int
__meta_track_apply(WT_SESSION_IMPL *session, WT_META_TRACK *trk, int unroll)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_RET;
int tret;
/*
* Unlock handles and complete checkpoints regardless of whether we are
* unrolling.
*/
if (!unroll && trk->op != WT_ST_CHECKPOINT && trk->op != WT_ST_LOCK)
goto free;
switch (trk->op) {
case WT_ST_EMPTY: /* Unused slot */
break;
case WT_ST_CHECKPOINT: /* Checkpoint, see above */
if (!unroll) {
btree = trk->dhandle->handle;
bm = btree->bm;
WT_WITH_DHANDLE(session, trk->dhandle,
WT_TRET(bm->checkpoint_resolve(bm, session)));
}
break;
case WT_ST_LOCK: /* Handle lock, see above */
if (unroll && trk->created)
F_SET(trk->dhandle, WT_DHANDLE_DISCARD);
WT_WITH_DHANDLE(session, trk->dhandle,
WT_TRET(__wt_session_release_btree(session)));
break;
case WT_ST_FILEOP: /* File operation */
/*
* For renames, both a and b are set.
* For creates, a is NULL.
* For removes, b is NULL.
*/
if (trk->a != NULL && trk->b != NULL &&
(tret = __wt_rename(session,
trk->b + strlen("file:"),
trk->a + strlen("file:"))) != 0) {
__wt_err(session, tret,
"metadata unroll rename %s to %s",
trk->b, trk->a);
WT_TRET(tret);
} else if (trk->a == NULL) {
if ((tret = __wt_remove(session,
trk->b + strlen("file:"))) != 0) {
__wt_err(session, tret,
"metadata unroll create %s",
trk->b);
WT_TRET(tret);
}
}
/*
* We can't undo removes yet: that would imply
* some kind of temporary rename and remove in
* roll forward.
*/
break;
case WT_ST_REMOVE: /* Remove trk.a */
if ((tret = __wt_metadata_remove(session, trk->a)) != 0) {
__wt_err(session, tret,
"metadata unroll remove: %s",
trk->a);
WT_TRET(tret);
}
break;
case WT_ST_SET: /* Set trk.a to trk.b */
if ((tret = __wt_metadata_update(
session, trk->a, trk->b)) != 0) {
__wt_err(session, tret,
"metadata unroll update %s to %s",
trk->a, trk->b);
WT_TRET(tret);
}
break;
WT_ILLEGAL_VALUE(session);
}
free: trk->op = WT_ST_EMPTY;
__wt_free(session, trk->a);
__wt_free(session, trk->b);
trk->dhandle = NULL;
return (ret);
}
/*
* __wt_session_compact --
* WT_SESSION.compact method.
*/
int
__wt_session_compact(
WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_COMPACT_STATE compact;
WT_CONFIG_ITEM cval;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
bool ignore_cache_size_set;
ignore_cache_size_set = false;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, compact, config, cfg);
/*
* The compaction thread should not block when the cache is full: it is
* holding locks blocking checkpoints and once the cache is full, it can
* spend a long time doing eviction.
*/
if (!F_ISSET(session, WT_SESSION_IGNORE_CACHE_SIZE)) {
ignore_cache_size_set = true;
F_SET(session, WT_SESSION_IGNORE_CACHE_SIZE);
}
/* In-memory ignores compaction operations. */
if (F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto err;
/*
* Non-LSM object compaction requires checkpoints, which are impossible
* in transactional contexts. Disallow in all contexts (there's no
* reason for LSM to allow this, possible or not), and check now so the
* error message isn't confusing.
*/
WT_ERR(__wt_txn_context_check(session, false));
/* Disallow objects in the WiredTiger name space. */
WT_ERR(__wt_str_name_check(session, uri));
if (!WT_PREFIX_MATCH(uri, "colgroup:") &&
!WT_PREFIX_MATCH(uri, "file:") &&
!WT_PREFIX_MATCH(uri, "index:") &&
!WT_PREFIX_MATCH(uri, "lsm:") &&
!WT_PREFIX_MATCH(uri, "table:")) {
if ((dsrc = __wt_schema_get_source(session, uri)) != NULL)
ret = dsrc->compact == NULL ?
__wt_object_unsupported(session, uri) :
dsrc->compact(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg);
else
ret = __wt_bad_object_type(session, uri);
goto err;
}
/* Setup the session handle's compaction state structure. */
memset(&compact, 0, sizeof(WT_COMPACT_STATE));
session->compact = &compact;
/* Compaction can be time-limited. */
WT_ERR(__wt_config_gets(session, cfg, "timeout", &cval));
session->compact->max_time = (uint64_t)cval.val;
__wt_epoch(session, &session->compact->begin);
/* Find the types of data sources being compacted. */
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, uri,
__compact_handle_append, __compact_uri_analyze, cfg, 0));
WT_ERR(ret);
if (session->compact->lsm_count != 0)
WT_ERR(__wt_schema_worker(
session, uri, NULL, __wt_lsm_compact, cfg, 0));
if (session->compact->file_count != 0)
WT_ERR(__compact_worker(session));
err: session->compact = NULL;
for (i = 0; i < session->op_handle_next; ++i) {
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__compact_end(session)));
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__wt_session_release_dhandle(session)));
}
__wt_free(session, session->op_handle);
session->op_handle_allocated = session->op_handle_next = 0;
/*
* Release common session resources (for example, checkpoint may acquire
* significant reconciliation structures/memory).
*/
WT_TRET(__wt_session_release_resources(session));
if (ignore_cache_size_set)
F_CLR(session, WT_SESSION_IGNORE_CACHE_SIZE);
if (ret != 0)
WT_STAT_CONN_INCR(session, session_table_compact_fail);
else
WT_STAT_CONN_INCR(session, session_table_compact_success);
API_END_RET_NOTFOUND_MAP(session, ret);
}
/*
* __compact_worker --
* Function to alternate between checkpoints and compaction calls.
*/
static int
__compact_worker(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
u_int i, loop;
bool another_pass;
/*
* Reset the handles' compaction skip flag (we don't bother setting
* or resetting it when we finish compaction, it's simpler to do it
* once, here).
*/
for (i = 0; i < session->op_handle_next; ++i)
session->op_handle[i]->compact_skip = false;
/*
* Perform an initial checkpoint (see this file's leading comment for
* details).
*/
WT_ERR(__compact_checkpoint(session));
/*
* We compact 10% of a file on each pass (but the overall size of the
* file is decreasing each time, so we're not compacting 10% of the
* original file each time). Try 100 times (which is clearly more than
* we need); quit if we make no progress.
*/
for (loop = 0; loop < 100; ++loop) {
/* Step through the list of files being compacted. */
for (another_pass = false,
i = 0; i < session->op_handle_next; ++i) {
/* Skip objects where there's no more work. */
if (session->op_handle[i]->compact_skip)
continue;
session->compact_state = WT_COMPACT_RUNNING;
WT_WITH_DHANDLE(session,
session->op_handle[i], ret = __wt_compact(session));
/*
* If successful and we did work, schedule another pass.
* If successful and we did no work, skip this file in
* the future.
*/
if (ret == 0) {
if (session->
compact_state == WT_COMPACT_SUCCESS)
another_pass = true;
else
session->
op_handle[i]->compact_skip = true;
continue;
}
/*
* If compaction failed because checkpoint was running,
* continue with the next handle. We might continue to
* race with checkpoint on each handle, but that's OK,
* we'll step through all the handles, and then we'll
* block until a checkpoint completes.
*
* Just quit if eviction is the problem.
*/
if (ret == EBUSY) {
if (__wt_cache_stuck(session)) {
WT_ERR_MSG(session, EBUSY,
"compaction halted by eviction "
"pressure");
}
ret = 0;
another_pass = true;
}
WT_ERR(ret);
}
if (!another_pass)
break;
/*
* Perform two checkpoints (see this file's leading comment for
* details).
*/
WT_ERR(__compact_checkpoint(session));
WT_ERR(__compact_checkpoint(session));
}
err: session->compact_state = WT_COMPACT_NONE;
return (ret);
}
/*
* __sweep --
* Close unused dhandles on the connection dhandle list.
*/
static int
__sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DATA_HANDLE *dhandle, *dhandle_next;
WT_DECL_RET;
time_t now;
int locked;
conn = S2C(session);
/* Don't discard handles that have been open recently. */
WT_RET(__wt_seconds(session, &now));
WT_STAT_FAST_CONN_INCR(session, dh_conn_sweeps);
dhandle = SLIST_FIRST(&conn->dhlh);
for (; dhandle != NULL; dhandle = dhandle_next) {
dhandle_next = SLIST_NEXT(dhandle, l);
if (WT_IS_METADATA(dhandle))
continue;
if (dhandle->session_inuse != 0 ||
now <= dhandle->timeofdeath + WT_DHANDLE_SWEEP_WAIT)
continue;
if (dhandle->timeofdeath == 0) {
dhandle->timeofdeath = now;
WT_STAT_FAST_CONN_INCR(session, dh_conn_tod);
continue;
}
/*
* We have a candidate for closing; if it's open, acquire an
* exclusive lock on the handle and close it. We might be
* blocking opens for a long time (over disk I/O), but the
* handle was quiescent for awhile.
*
* The close can fail if an update cannot be written (updates
* in a no-longer-referenced file might not yet be globally
* visible if sessions have disjoint sets of files open). If
* the handle is busy, skip it, we'll retry the close the next
* time, after the transaction state has progressed.
*
* We don't set WT_DHANDLE_EXCLUSIVE deliberately, we want
* opens to block on us rather than returning an EBUSY error to
* the application.
*/
if ((ret =
__wt_try_writelock(session, dhandle->rwlock)) == EBUSY)
continue;
WT_RET(ret);
locked = 1;
/* If the handle is open, try to close it. */
if (F_ISSET(dhandle, WT_DHANDLE_OPEN)) {
WT_WITH_DHANDLE(session, dhandle,
ret = __wt_conn_btree_sync_and_close(session, 0));
if (ret != 0)
goto unlock;
/* We closed the btree handle, bump the statistic. */
WT_STAT_FAST_CONN_INCR(session, dh_conn_handles);
}
/*
* If there are no longer any references to the handle in any
* sessions, attempt to discard it. The called function
* re-checks that the handle is not in use, which is why we
* don't do any special handling of EBUSY returns above.
*/
if (dhandle->session_inuse == 0 && dhandle->session_ref == 0) {
WT_WITH_DHANDLE(session, dhandle,
ret = __wt_conn_dhandle_discard_single(session, 0));
if (ret != 0)
goto unlock;
/* If the handle was discarded, it isn't locked. */
locked = 0;
} else
WT_STAT_FAST_CONN_INCR(session, dh_conn_ref);
unlock: if (locked)
WT_TRET(__wt_writeunlock(session, dhandle->rwlock));
WT_RET_BUSY_OK(ret);
}
return (0);
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_SESSION_IMPL *ckpt_session;
int saved_ret;
bool did_drop;
saved_ret = 0;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, skip unnecessary metadata
* checkpoints. For example, this happens if attempting to create a
* data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
goto err;
/* Unrolling doesn't require syncing the metadata. */
if (unroll)
goto err;
if (F_ISSET(session, WT_SESSION_SCHEMA_TXN)) {
F_CLR(session, WT_SESSION_SCHEMA_TXN);
#ifdef WT_ENABLE_SCHEMA_TXN
WT_ERR(__wt_txn_commit(session, NULL));
__wt_errx(session, "TRACK: Commit internal schema txn");
#endif
}
/*
* If we don't have the metadata cursor (e.g, we're in the process of
* creating the metadata), we can't sync it.
*/
if (!need_sync || session->meta_cursor == NULL ||
F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto err;
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED))
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_txn_checkpoint_log(
session, false, WT_TXN_LOG_CKPT_SYNC, NULL));
else {
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
ckpt_session = S2C(session)->meta_ckpt_session;
/*
* If this operation is part of a running transaction, that
* should be included in the checkpoint.
*/
ckpt_session->txn.id = session->txn.id;
WT_ASSERT(session,
!F_ISSET(session, WT_SESSION_LOCKED_METADATA));
WT_WITH_DHANDLE(ckpt_session, WT_SESSION_META_DHANDLE(session),
WT_WITH_METADATA_LOCK(ckpt_session,
ret = __wt_checkpoint(ckpt_session, NULL)));
ckpt_session->txn.id = WT_TXN_NONE;
if (ret == 0)
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint_sync(session, NULL));
}
err: /*
* Undo any tracked operations on failure.
* Apply any tracked operations post-commit.
*/
did_drop = false;
if (unroll || ret != 0) {
saved_ret = ret;
ret = 0;
while (--trk >= trk_orig) {
did_drop = did_drop || trk->op == WT_ST_DROP_COMMIT;
WT_TRET(__meta_track_unroll(session, trk));
}
} else
for (; trk_orig < trk; trk_orig++) {
did_drop = did_drop ||
trk_orig->op == WT_ST_DROP_COMMIT;
WT_TRET(__meta_track_apply(session, trk_orig));
}
if (F_ISSET(session, WT_SESSION_SCHEMA_TXN)) {
F_CLR(session, WT_SESSION_SCHEMA_TXN);
/*
* We should have committed above unless we're unrolling, there
* was an error or the operation was a noop.
*/
WT_ASSERT(session, unroll || saved_ret != 0 ||
session->txn.mod_count == 0);
#ifdef WT_ENABLE_SCHEMA_TXN
__wt_err(session, saved_ret,
"TRACK: Abort internal schema txn");
WT_TRET(__wt_txn_rollback(session, NULL));
#endif
}
/*
* Wake up the sweep thread: particularly for the in-memory
* storage engine, we want to reclaim space immediately.
*/
if (did_drop && S2C(session)->sweep_cond != NULL)
__wt_cond_signal(session, S2C(session)->sweep_cond);
if (ret != 0)
WT_PANIC_RET(session, ret,
"failed to apply or unroll all tracked operations");
return (saved_ret == 0 ? 0 : saved_ret);
}
/*
* __drop_table --
* WT_SESSION::drop for a table.
*/
static int
__drop_table(
WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_COLGROUP *colgroup;
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
u_int i;
const char *name;
bool tracked;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_TABLE_WRITE));
name = uri;
WT_PREFIX_SKIP_REQUIRED(session, name, "table:");
table = NULL;
tracked = false;
/*
* Open the table so we can drop its column groups and indexes.
*
* Ideally we would keep the table locked exclusive across the drop,
* but for now we rely on the global table lock to prevent the table
* being reopened while it is being dropped. One issue is that the
* WT_WITHOUT_LOCKS macro can drop and reacquire the global table lock,
* avoiding deadlocks while waiting for LSM operation to quiesce.
*
* Temporarily getting the table exclusively serves the purpose
* of ensuring that cursors on the table that are already open
* must at least be closed before this call proceeds.
*/
WT_ERR(__wt_schema_get_table_uri(session, uri, true,
WT_DHANDLE_EXCLUSIVE, &table));
WT_ERR(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(session, uri, true, 0, &table));
/* Drop the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if ((colgroup = table->cgroups[i]) == NULL)
continue;
/*
* Drop the column group before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, colgroup->source, cfg));
WT_ERR(__wt_metadata_remove(session, colgroup->name));
}
/* Drop the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
if ((idx = table->indices[i]) == NULL)
continue;
/*
* Drop the index before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, idx->source, cfg));
WT_ERR(__wt_metadata_remove(session, idx->name));
}
/* Make sure the table data handle is closed. */
WT_TRET(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(
session, uri, true, WT_DHANDLE_EXCLUSIVE, &table));
F_SET(&table->iface, WT_DHANDLE_DISCARD);
if (WT_META_TRACKING(session)) {
WT_WITH_DHANDLE(session, &table->iface,
ret = __wt_meta_track_handle_lock(session, false));
WT_ERR(ret);
tracked = true;
}
/* Remove the metadata entry (ignore missing items). */
WT_ERR(__wt_metadata_remove(session, uri));
err: if (table != NULL && !tracked)
WT_TRET(__wt_schema_release_table(session, table));
return (ret);
}
