/*
* __wt_lsm_tree_drop --
* Drop an LSM tree.
*/
int
__wt_lsm_tree_drop(
WT_SESSION_IMPL *session, const char *name, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
u_int i;
int locked;
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, name, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_ERR(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_ERR(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Drop the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
WT_ERR(__wt_schema_drop(session, chunk->uri, cfg));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(
__wt_schema_drop(session, chunk->bloom_uri, cfg));
}
/* Drop any chunks on the obsolete list. */
for (i = 0; i < lsm_tree->nold_chunks; i++) {
if ((chunk = lsm_tree->old_chunks[i]) == NULL)
continue;
WT_ERR(__wt_schema_drop(session, chunk->uri, cfg));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(
__wt_schema_drop(session, chunk->bloom_uri, cfg));
}
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
ret = __wt_metadata_remove(session, name);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __wt_lsm_tree_truncate --
* Truncate an LSM tree.
*/
int
__wt_lsm_tree_truncate(
WT_SESSION_IMPL *session, const char *name, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
int locked;
WT_UNUSED(cfg);
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, name, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_RET(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_RET(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Create the new chunk. */
WT_ERR(__wt_calloc_def(session, 1, &chunk));
chunk->id = WT_ATOMIC_ADD(lsm_tree->last, 1);
WT_ERR(__wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
/* Mark all chunks old. */
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, 0, lsm_tree->nchunks, chunk));
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
WT_ERR(__lsm_tree_start_worker(session, lsm_tree));
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
/*
* Don't discard the LSM tree structure unless there has been an
* error. The handle remains valid for future operations.
*/
if (ret != 0)
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __log_server --
* The log server thread.
*/
static WT_THREAD_RET
__log_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_SESSION_IMPL *session;
u_int locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = 0;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Perform log pre-allocation.
*/
if (conn->log_prealloc > 0)
WT_ERR(__log_prealloc_once(session));
/*
* Perform the archive.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ARCHIVE)) {
if (__wt_try_writelock(
session, log->log_archive_lock) == 0) {
locked = 1;
WT_ERR(__log_archive_once(session, 0));
WT_ERR( __wt_writeunlock(
session, log->log_archive_lock));
locked = 0;
} else
WT_ERR(__wt_verbose(session, WT_VERB_LOG,
"log_archive: Blocked due to open log "
"cursor holding archive lock"));
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait(session, conn->log_cond, WT_MILLION));
}
if (0) {
err: __wt_err(session, ret, "log server error");
}
if (locked)
(void)__wt_writeunlock(session, log->log_archive_lock);
return (WT_THREAD_RET_VALUE);
}
/*
* __wt_thread_group_start_one --
* Start a new thread if possible.
*/
int
__wt_thread_group_start_one(
WT_SESSION_IMPL *session, WT_THREAD_GROUP *group, bool wait)
{
WT_DECL_RET;
if (group->current_threads >= group->max)
return (0);
if (wait)
__wt_writelock(session, group->lock);
else if (__wt_try_writelock(session, group->lock) != 0)
return (0);
/* Recheck the bounds now that we hold the lock */
if (group->current_threads < group->max)
WT_TRET(__thread_group_grow(
session, group, group->current_threads + 1));
__wt_writeunlock(session, group->lock);
return (ret);
}
/*
* __log_server --
* The log server thread.
*/
static WT_THREAD_RET
__log_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_SESSION_IMPL *session;
int freq_per_sec, signalled;
session = arg;
conn = S2C(session);
log = conn->log;
signalled = 0;
/*
* Set this to the number of times per second we want to force out the
* log slot buffer.
*/
#define WT_FORCE_PER_SECOND 20
freq_per_sec = WT_FORCE_PER_SECOND;
/*
* The log server thread does a variety of work. It forces out any
* buffered log writes. It pre-allocates log files and it performs
* log archiving. The reason the wrlsn thread does not force out
* the buffered writes is because we want to process and move the
* write_lsn forward as quickly as possible. The same reason applies
* to why the log file server thread does not force out the writes.
* That thread does fsync calls which can take a long time and we
* don't want log records sitting in the buffer over the time it
* takes to sync out an earlier file.
*/
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Slots depend on future activity. Force out buffered
* writes in case we are idle. This cannot be part of the
* wrlsn thread because of interaction advancing the write_lsn
* and a buffer may need to wait for the write_lsn to advance
* in the case of a synchronous buffer. We end up with a hang.
*/
WT_ERR_BUSY_OK(__wt_log_force_write(session, 0));
/*
* We don't want to archive or pre-allocate files as often as
* we want to force out log buffers. Only do it once per second
* or if the condition was signalled.
*/
if (--freq_per_sec <= 0 || signalled != 0) {
freq_per_sec = WT_FORCE_PER_SECOND;
/*
* Perform log pre-allocation.
*/
if (conn->log_prealloc > 0)
WT_ERR(__log_prealloc_once(session));
/*
* Perform the archive.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ARCHIVE)) {
if (__wt_try_writelock(
session, log->log_archive_lock) == 0) {
ret = __log_archive_once(session, 0);
WT_TRET(__wt_writeunlock(
session, log->log_archive_lock));
WT_ERR(ret);
} else
WT_ERR(
__wt_verbose(session, WT_VERB_LOG,
"log_archive: Blocked due to open "
"log cursor holding archive lock"));
}
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait_signal(session, conn->log_cond,
WT_MILLION / WT_FORCE_PER_SECOND, &signalled));
}
if (0) {
err: __wt_err(session, ret, "log server error");
}
return (WT_THREAD_RET_VALUE);
}
/*
* __wt_lsm_tree_rename --
* Rename an LSM tree.
*/
int
__wt_lsm_tree_rename(WT_SESSION_IMPL *session,
const char *oldname, const char *newname, const char *cfg[])
{
WT_DECL_RET;
WT_ITEM buf;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
const char *old;
u_int i;
int locked;
old = NULL;
WT_CLEAR(buf);
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, oldname, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_ERR(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_ERR(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Set the new name. */
WT_ERR(__lsm_tree_set_name(session, lsm_tree, newname));
/* Rename the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
old = chunk->uri;
chunk->uri = NULL;
WT_ERR(__wt_lsm_tree_chunk_name(
session, lsm_tree, chunk->id, &buf));
chunk->uri = __wt_buf_steal(session, &buf, NULL);
WT_ERR(__wt_schema_rename(session, old, chunk->uri, cfg));
__wt_free(session, old);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
old = chunk->bloom_uri;
chunk->bloom_uri = NULL;
WT_ERR(__wt_lsm_tree_bloom_name(
session, lsm_tree, chunk->id, &buf));
chunk->bloom_uri = __wt_buf_steal(session, &buf, NULL);
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
WT_ERR(__wt_schema_rename(
session, old, chunk->uri, cfg));
__wt_free(session, old);
}
}
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
ret = __wt_lsm_meta_write(session, lsm_tree);
if (ret == 0)
ret = __wt_metadata_remove(session, oldname);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
if (old != NULL)
__wt_free(session, old);
/*
* Discard this LSM tree structure. The first operation on the renamed
* tree will create a new one.
*/
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __log_server --
* The log server thread.
*/
static WT_THREAD_RET
__log_server(void *arg)
{
struct timespec start, now;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_SESSION_IMPL *session;
uint64_t timediff;
bool did_work, locked, signalled;
session = arg;
conn = S2C(session);
log = conn->log;
locked = signalled = false;
/*
* Set this to the number of milliseconds we want to run archive and
* pre-allocation. Start it so that we run on the first time through.
*/
timediff = WT_THOUSAND;
/*
* The log server thread does a variety of work. It forces out any
* buffered log writes. It pre-allocates log files and it performs
* log archiving. The reason the wrlsn thread does not force out
* the buffered writes is because we want to process and move the
* write_lsn forward as quickly as possible. The same reason applies
* to why the log file server thread does not force out the writes.
* That thread does fsync calls which can take a long time and we
* don't want log records sitting in the buffer over the time it
* takes to sync out an earlier file.
*/
did_work = true;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Slots depend on future activity. Force out buffered
* writes in case we are idle. This cannot be part of the
* wrlsn thread because of interaction advancing the write_lsn
* and a buffer may need to wait for the write_lsn to advance
* in the case of a synchronous buffer. We end up with a hang.
*/
WT_ERR_BUSY_OK(__wt_log_force_write(session, 0, &did_work));
/*
* We don't want to archive or pre-allocate files as often as
* we want to force out log buffers. Only do it once per second
* or if the condition was signalled.
*/
if (timediff >= WT_THOUSAND || signalled) {
/*
* Perform log pre-allocation.
*/
if (conn->log_prealloc > 0) {
/*
* Log file pre-allocation is disabled when a
* hot backup cursor is open because we have
* agreed not to rename or remove any files in
* the database directory.
*/
WT_ERR(__wt_readlock(
session, conn->hot_backup_lock));
locked = true;
if (!conn->hot_backup)
WT_ERR(__log_prealloc_once(session));
WT_ERR(__wt_readunlock(
session, conn->hot_backup_lock));
locked = false;
}
/*
* Perform the archive.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ARCHIVE)) {
if (__wt_try_writelock(
session, log->log_archive_lock) == 0) {
ret = __log_archive_once(session, 0);
WT_TRET(__wt_writeunlock(
session, log->log_archive_lock));
WT_ERR(ret);
} else
WT_ERR(
__wt_verbose(session, WT_VERB_LOG,
"log_archive: Blocked due to open "
"log cursor holding archive lock"));
}
}
/* Wait until the next event. */
WT_ERR(__wt_epoch(session, &start));
WT_ERR(__wt_cond_auto_wait_signal(session, conn->log_cond,
did_work, &signalled));
WT_ERR(__wt_epoch(session, &now));
timediff = WT_TIMEDIFF_MS(now, start);
}
if (0) {
err: __wt_err(session, ret, "log server error");
if (locked)
WT_TRET(__wt_readunlock(
session, conn->hot_backup_lock));
}
return (WT_THREAD_RET_VALUE);
}
/*
* __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);
}
/*
* __wt_txn_update_oldest --
* Sweep the running transactions to update the oldest ID required.
*/
int
__wt_txn_update_oldest(WT_SESSION_IMPL *session, uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_SESSION_IMPL *oldest_session;
WT_TXN_GLOBAL *txn_global;
uint64_t current_id, last_running, oldest_id;
uint64_t prev_last_running, prev_oldest_id;
bool strict, wait;
conn = S2C(session);
txn_global = &conn->txn_global;
strict = LF_ISSET(WT_TXN_OLDEST_STRICT);
wait = LF_ISSET(WT_TXN_OLDEST_WAIT);
current_id = last_running = txn_global->current;
prev_last_running = txn_global->last_running;
prev_oldest_id = txn_global->oldest_id;
/*
* For pure read-only workloads, or if the update isn't forced and the
* oldest ID isn't too far behind, avoid scanning.
*/
if (prev_oldest_id == current_id ||
(!strict && WT_TXNID_LT(current_id, prev_oldest_id + 100)))
return (0);
/* First do a read-only scan. */
if (wait)
__wt_readlock(session, txn_global->scan_rwlock);
else if ((ret =
__wt_try_readlock(session, txn_global->scan_rwlock)) != 0)
return (ret == EBUSY ? 0 : ret);
__txn_oldest_scan(session, &oldest_id, &last_running, &oldest_session);
__wt_readunlock(session, txn_global->scan_rwlock);
/*
* If the state hasn't changed (or hasn't moved far enough for
* non-forced updates), give up.
*/
if ((oldest_id == prev_oldest_id ||
(!strict && WT_TXNID_LT(oldest_id, prev_oldest_id + 100))) &&
((last_running == prev_last_running) ||
(!strict && WT_TXNID_LT(last_running, prev_last_running + 100))))
return (0);
/* It looks like an update is necessary, wait for exclusive access. */
if (wait)
__wt_writelock(session, txn_global->scan_rwlock);
else if ((ret =
__wt_try_writelock(session, txn_global->scan_rwlock)) != 0)
return (ret == EBUSY ? 0 : ret);
/*
* If the oldest ID has been updated while we waited, don't bother
* scanning.
*/
if (WT_TXNID_LE(oldest_id, txn_global->oldest_id) &&
WT_TXNID_LE(last_running, txn_global->last_running))
goto done;
/*
* Re-scan now that we have exclusive access. This is necessary because
* threads get transaction snapshots with read locks, and we have to be
* sure that there isn't a thread that has got a snapshot locally but
* not yet published its snap_min.
*/
__txn_oldest_scan(session, &oldest_id, &last_running, &oldest_session);
#ifdef HAVE_DIAGNOSTIC
{
/*
* Make sure the ID doesn't move past any named snapshots.
*
* Don't include the read/assignment in the assert statement. Coverity
* complains if there are assignments only done in diagnostic builds,
* and when the read is from a volatile.
*/
uint64_t id = txn_global->nsnap_oldest_id;
WT_ASSERT(session,
id == WT_TXN_NONE || !WT_TXNID_LT(id, oldest_id));
}
#endif
/* Update the oldest ID. */
if (WT_TXNID_LT(txn_global->oldest_id, oldest_id))
txn_global->oldest_id = oldest_id;
if (WT_TXNID_LT(txn_global->last_running, last_running)) {
txn_global->last_running = last_running;
#ifdef HAVE_VERBOSE
/* Output a verbose message about long-running transactions,
* but only when some progress is being made. */
if (WT_VERBOSE_ISSET(session, WT_VERB_TRANSACTION) &&
current_id - oldest_id > 10000 && oldest_session != NULL) {
__wt_verbose(session, WT_VERB_TRANSACTION,
"old snapshot %" PRIu64
" pinned in session %" PRIu32 " [%s]"
" with snap_min %" PRIu64 "\n",
oldest_id, oldest_session->id,
oldest_session->lastop,
oldest_session->txn.snap_min);
}
#endif
}
done: __wt_writeunlock(session, txn_global->scan_rwlock);
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);
}
/*
* __conn_dhandle_open_lock --
* Spin on the current data handle until either (a) it is open, read
* locked; or (b) it is closed, write locked. If exclusive access is
* requested and cannot be granted immediately because the handle is
* in use, fail with EBUSY.
*
* Here is a brief summary of how different operations synchronize using
* either the schema lock, handle locks or handle flags:
*
* open -- holds the schema lock, one thread gets the handle exclusive,
* reverts to a shared handle lock and drops the schema lock
* once the handle is open;
* bulk load -- sets bulk and exclusive;
* salvage, truncate, update, verify -- hold the schema lock, set a
* "special" flag;
* sweep -- gets a write lock on the handle, doesn't set exclusive
*
* The schema lock prevents a lot of potential conflicts: we should never
* see handles being salvaged or verified because those operation hold the
* schema lock. However, it is possible to see a handle that is being
* bulk loaded, or that the sweep server is closing.
*
* The principle here is that application operations can cause other
* application operations to fail (so attempting to open a cursor on a
* file while it is being bulk-loaded will fail), but internal or
* database-wide operations should not prevent application-initiated
* operations. For example, attempting to verify a file should not fail
* because the sweep server happens to be in the process of closing that
* file.
*/
static int
__conn_dhandle_open_lock(
WT_SESSION_IMPL *session, WT_DATA_HANDLE *dhandle, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
int is_open, lock_busy, want_exclusive;
btree = dhandle->handle;
lock_busy = 0;
want_exclusive = LF_ISSET(WT_DHANDLE_EXCLUSIVE) ? 1 : 0;
/*
* Check that the handle is open. We've already incremented
* the reference count, so once the handle is open it won't be
* closed by another thread.
*
* If we can see the WT_DHANDLE_OPEN flag set while holding a
* lock on the handle, then it's really open and we can start
* using it. Alternatively, if we can get an exclusive lock
* and WT_DHANDLE_OPEN is still not set, we need to do the open.
*/
for (;;) {
/*
* If the handle is already open for a special operation,
* give up.
*/
if (F_ISSET(btree, WT_BTREE_SPECIAL_FLAGS))
return (EBUSY);
/*
* If the handle is open, get a read lock and recheck.
*
* Wait for a read lock if we want exclusive access and failed
* to get it: the sweep server may be closing this handle, and
* we need to wait for it to complete. If we want exclusive
* access and find the handle open once we get the read lock,
* give up: some other thread has it locked for real.
*/
if (F_ISSET(dhandle, WT_DHANDLE_OPEN) &&
(!want_exclusive || lock_busy)) {
WT_RET(__wt_readlock(session, dhandle->rwlock));
is_open = F_ISSET(dhandle, WT_DHANDLE_OPEN) ? 1 : 0;
if (is_open && !want_exclusive)
return (0);
WT_RET(__wt_readunlock(session, dhandle->rwlock));
} else
is_open = 0;
/*
* It isn't open or we want it exclusive: try to get an
* exclusive lock. There is some subtlety here: if we race
* with another thread that successfully opens the file, we
* don't want to block waiting to get exclusive access.
*/
if ((ret = __wt_try_writelock(session, dhandle->rwlock)) == 0) {
/*
* If it was opened while we waited, drop the write
* lock and get a read lock instead.
*/
if (F_ISSET(dhandle, WT_DHANDLE_OPEN) &&
!want_exclusive) {
lock_busy = 0;
WT_RET(
__wt_writeunlock(session, dhandle->rwlock));
continue;
}
/* We have an exclusive lock, we're done. */
F_SET(dhandle, WT_DHANDLE_EXCLUSIVE);
return (0);
} else if (ret != EBUSY || (is_open && want_exclusive))
return (ret);
else
lock_busy = 1;
/* Give other threads a chance to make progress. */
__wt_yield();
}
}