/*
* __wt_clsm_await_switch --
* Wait for a switch to have completed in the LSM tree
*/
int
__wt_clsm_await_switch(WT_CURSOR_LSM *clsm)
{
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
int waited;
lsm_tree = clsm->lsm_tree;
session = (WT_SESSION_IMPL *)clsm->iface.session;
/*
* If there is no primary chunk, or a chunk has overflowed the hard
* limit, which either means a worker thread has fallen behind or there
* has just been a user-level checkpoint, wait until the tree changes.
*
* We used to switch chunks in the application thread here, but that is
* problematic because there is a transaction in progress and it could
* roll back, leaving the metadata inconsistent.
*/
for (waited = 0;
lsm_tree->nchunks == 0 ||
clsm->dsk_gen == lsm_tree->dsk_gen;
++waited) {
if (waited % 1000 == 0)
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_SWITCH, 0, lsm_tree));
__wt_sleep(0, 10);
}
return (0);
}
/*
* __wt_attach --
* A routine to wait for the debugging to attach.
*/
void
__wt_attach(WT_SESSION_IMPL *session)
{
#ifdef HAVE_ATTACH
__wt_errx(session, "process ID %" PRIdMAX
": waiting for debugger...", (intmax_t)getpid());
/* Sleep forever, the debugger will interrupt us when it attaches. */
for (;;)
__wt_sleep(100, 0);
#else
WT_UNUSED(session);
#endif
}
/*
* compaction --
* Periodically do a compaction operation.
*/
WT_THREAD_RET
compact(void *arg)
{
WT_CONNECTION *conn;
WT_DECL_RET;
WT_SESSION *session;
u_int period;
(void)(arg);
/* Compaction isn't supported for all data sources. */
if (DATASOURCE("helium") || DATASOURCE("kvsbdb"))
return (WT_THREAD_RET_VALUE);
/* Open a session. */
conn = g.wts_conn;
testutil_check(conn->open_session(conn, NULL, NULL, &session));
/*
* Perform compaction at somewhere under 15 seconds (so we get at
* least one done), and then at 23 second intervals.
*/
for (period = mmrand(NULL, 1, 15);; period = 23) {
/* Sleep for short periods so we don't make the run wait. */
while (period > 0 && !g.workers_finished) {
--period;
__wt_sleep(1, 0);
}
if (g.workers_finished)
break;
/*
* Compact can return EBUSY if concurrent with alter or if there
* is eviction pressure, or we collide with checkpoints.
*
* Compact returns ETIMEDOUT if the compaction doesn't finish in
* in some number of seconds. We don't configure a timeout and
* occasionally exceed the default of 1200 seconds.
*/
ret = session->compact(session, g.uri, NULL);
if (ret != 0 &&
ret != EBUSY && ret != ETIMEDOUT && ret != WT_ROLLBACK)
testutil_die(ret, "session.compact");
}
testutil_check(session->close(session, NULL));
return (WT_THREAD_RET_VALUE);
}
/*
* __wt_log_slot_wait --
* Wait for slot leader to allocate log area and tell us our log offset.
*/
int
__wt_log_slot_wait(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
{
int yield_count;
yield_count = 0;
WT_UNUSED(session);
while (slot->slot_state > WT_LOG_SLOT_DONE)
if (++yield_count < 1000)
__wt_yield();
else
__wt_sleep(0, 200);
return (0);
}
/*
* __wt_lsm_compact --
* Compact an LSM tree called via __wt_schema_worker.
*/
int
__wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, int *skip)
{
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
uint64_t last_merge_progressing;
time_t begin, end;
/*
* This function is applied to all matching sources: ignore anything
* that is not an LSM tree.
*/
if (!WT_PREFIX_MATCH(name, "lsm:"))
return (0);
/* Tell __wt_schema_worker not to look inside the LSM tree. */
*skip = 1;
WT_RET(__wt_lsm_tree_get(session, name, 0, &lsm_tree));
if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE) ||
lsm_tree->merge_threads == 0)
WT_RET_MSG(session, EINVAL,
"LSM compaction requires active merge threads");
WT_RET(__wt_seconds(session, &begin));
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
/* Wake up the merge threads. */
WT_RET(__wt_cond_signal(session, lsm_tree->work_cond));
/* Now wait for merge activity to stop. */
do {
last_merge_progressing = lsm_tree->merge_progressing;
__wt_sleep(1, 0);
WT_RET(__wt_seconds(session, &end));
if (session->compact->max_time > 0 &&
session->compact->max_time < (uint64_t)(end - begin))
WT_ERR(ETIMEDOUT);
} while (lsm_tree->merge_progressing != last_merge_progressing &&
lsm_tree->nchunks > 1);
err: F_CLR(lsm_tree, WT_LSM_TREE_COMPACTING);
return (ret);
}
/*
* __compact_checkpoint --
* Perform a checkpoint for compaction.
*/
static int
__compact_checkpoint(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
WT_TXN_GLOBAL *txn_global;
uint64_t txn_gen;
/*
* Force compaction checkpoints: we don't want to skip it because the
* work we need to have done is done in the underlying block manager.
*/
const char *checkpoint_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_checkpoint), "force=1", NULL };
/* Checkpoints take a lot of time, check if we've run out. */
WT_RET(__wt_session_compact_check_timeout(session));
if ((ret = __wt_txn_checkpoint(session, checkpoint_cfg, false)) == 0)
return (0);
WT_RET_BUSY_OK(ret);
/*
* If there's a checkpoint running, wait for it to complete, checking if
* we're out of time. If there's no checkpoint running or the checkpoint
* generation number changes, the checkpoint blocking us has completed.
*/
txn_global = &S2C(session)->txn_global;
for (txn_gen = __wt_gen(session, WT_GEN_CHECKPOINT);;) {
/*
* This loop only checks objects that are declared volatile,
* therefore no barriers are needed.
*/
if (!txn_global->checkpoint_running ||
txn_gen != __wt_gen(session, WT_GEN_CHECKPOINT))
break;
WT_RET(__wt_session_compact_check_timeout(session));
__wt_sleep(2, 0);
}
return (0);
}
/*
* thread_ts_run --
* Runner function for a timestamp thread.
*/
static WT_THREAD_RET
thread_ts_run(void *arg)
{
WT_DECL_RET;
WT_SESSION *session;
THREAD_DATA *td;
char tscfg[64], ts_string[WT_TS_HEX_STRING_SIZE];
td = (THREAD_DATA *)arg;
testutil_check(td->conn->open_session(td->conn, NULL, NULL, &session));
/* Update the oldest timestamp every 1 millisecond. */
for (;;) {
/*
* We get the last committed timestamp periodically in order to
* update the oldest timestamp, that requires locking out
* transactional ops that set or query a timestamp.
*/
testutil_check(pthread_rwlock_wrlock(&ts_lock));
ret = td->conn->query_timestamp(
td->conn, ts_string, "get=all_committed");
testutil_check(pthread_rwlock_unlock(&ts_lock));
testutil_assert(ret == 0 || ret == WT_NOTFOUND);
if (ret == 0) {
/*
* Set both the oldest and stable timestamp so that we
* don't need to maintain read availability at older
* timestamps.
*/
testutil_check(__wt_snprintf(
tscfg, sizeof(tscfg),
"oldest_timestamp=%s,stable_timestamp=%s",
ts_string, ts_string));
testutil_check(
td->conn->set_timestamp(td->conn, tscfg));
}
__wt_sleep(0, 1000);
}
/* NOTREACHED */
}
/*
* __logmgr_force_ckpt --
* Force a checkpoint out, waiting for the checkpoint LSN in the log
* is up to the given log number.
*/
static int
__logmgr_force_ckpt(WT_SESSION_IMPL *session, uint32_t lognum)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_SESSION_IMPL *tmp_session;
int yield;
conn = S2C(session);
log = conn->log;
yield = 0;
WT_RET(__wt_open_internal_session(conn,
"compatibility-reconfig", true, 0, &tmp_session));
while (log->ckpt_lsn.l.file < lognum) {
/*
* Force a checkpoint to be written in the new log file and
* force the archiving of all previous log files. We do the
* checkpoint in the loop because the checkpoint LSN in the
* log record could still reflect the previous log file in
* cases such as the write LSN has not yet advanced into the
* new log file due to another group of threads still in
* progress with their slot copies or writes.
*/
WT_RET(tmp_session->iface.checkpoint(
&tmp_session->iface, "force=1"));
WT_RET(WT_SESSION_CHECK_PANIC(tmp_session));
/*
* Only sleep in the rare case that we had to come through
* this loop more than once.
*/
if (yield++) {
WT_STAT_CONN_INCR(session, log_force_ckpt_sleep);
__wt_sleep(0, WT_THOUSAND);
}
}
WT_RET(tmp_session->iface.close(&tmp_session->iface, NULL));
return (0);
}
/*
* __wt_ref_out --
* Discard an in-memory page, freeing all memory associated with it.
*/
void
__wt_ref_out(WT_SESSION_IMPL *session, WT_REF *ref)
{
/*
* A version of the page-out function that allows us to make additional
* diagnostic checks.
*
* The WT_REF cannot be the eviction thread's location.
*/
WT_ASSERT(session, S2BT(session)->evict_ref != ref);
#ifdef HAVE_DIAGNOSTIC
{
WT_HAZARD *hp;
int i;
/*
* Make sure no other thread has a hazard pointer on the page we are
* about to discard. This is complicated by the fact that readers
* publish their hazard pointer before re-checking the page state, so
* our check can race with readers without indicating a real problem.
* Wait for up to a second for hazard pointers to be cleared.
*/
for (hp = NULL, i = 0; i < 100; i++) {
if ((hp = __wt_hazard_check(session, ref)) == NULL)
break;
__wt_sleep(0, 10000);
}
if (hp != NULL)
__wt_errx(session,
"discarded page has hazard pointer: (%p: %s, line %d)",
(void *)hp->ref, hp->file, hp->line);
WT_ASSERT(session, hp == NULL);
}
#endif
__wt_page_out(session, &ref->page);
}
/*
* __wt_lsm_compact --
* Compact an LSM tree called via __wt_schema_worker.
*/
int
__wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, int *skip)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
time_t begin, end;
uint64_t progress;
int i, compacting, flushing, locked, ref;
compacting = flushing = locked = ref = 0;
chunk = NULL;
/*
* This function is applied to all matching sources: ignore anything
* that is not an LSM tree.
*/
if (!WT_PREFIX_MATCH(name, "lsm:"))
return (0);
/* Tell __wt_schema_worker not to look inside the LSM tree. */
*skip = 1;
WT_RET(__wt_lsm_tree_get(session, name, 0, &lsm_tree));
if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE))
WT_ERR_MSG(session, EINVAL,
"LSM compaction requires active merge threads");
WT_ERR(__wt_seconds(session, &begin));
/*
* Compacting has two distinct phases.
* 1. All in-memory chunks up to and including the current
* current chunk must be flushed. Normally, the flush code
* does not flush the last, in-use chunk, so we set a force
* flag to include that last chunk. We monitor the state of the
* last chunk and periodically push another forced flush work
* unit until it is complete.
* 2. After all flushing is done, we move onto the merging
* phase for compaction. Again, we monitor the state and
* continue to push merge work units until all merging is done.
*/
/* Lock the tree: single-thread compaction. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = 1;
/* Clear any merge throttle: compact throws out that calculation. */
lsm_tree->merge_throttle = 0;
lsm_tree->merge_aggressiveness = 0;
progress = lsm_tree->merge_progressing;
/* If another thread started a compact on this tree, we're done. */
if (F_ISSET(lsm_tree, WT_LSM_TREE_COMPACTING))
goto err;
/*
* Set the switch transaction on the current chunk, if it
* hasn't been set before. This prevents further writes, so it
* can be flushed by the checkpoint worker.
*/
if (lsm_tree->nchunks > 0 &&
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) != NULL) {
if (chunk->switch_txn == WT_TXN_NONE)
chunk->switch_txn = __wt_txn_new_id(session);
/*
* If we have a chunk, we want to look for it to be on-disk.
* So we need to add a reference to keep it available.
*/
(void)WT_ATOMIC_ADD4(chunk->refcnt, 1);
ref = 1;
}
locked = 0;
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (chunk != NULL) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Compact force flush %s flags 0x%" PRIx32
" chunk %u flags 0x%"
PRIx32, name, lsm_tree->flags, chunk->id, chunk->flags));
flushing = 1;
/*
* Make sure the in-memory chunk gets flushed do not push a
* switch, because we don't want to create a new in-memory
* chunk if the tree is being used read-only now.
*/
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_FLUSH, WT_LSM_WORK_FORCE, lsm_tree));
} else {
/*
* If there is no chunk to flush, go straight to the
* compacting state.
*/
compacting = 1;
progress = lsm_tree->merge_progressing;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"COMPACT: Start compacting %s", lsm_tree->name));
}
/* Wait for the work unit queues to drain. */
while (F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE)) {
/*
* The flush flag is cleared when the chunk has been flushed.
* Continue to push forced flushes until the chunk is on disk.
* Once it is on disk move to the compacting phase.
*/
if (flushing) {
WT_ASSERT(session, chunk != NULL);
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
WT_ERR(__wt_verbose(session,
WT_VERB_LSM,
"Compact flush done %s chunk %u. "
"Start compacting progress %" PRIu64,
name, chunk->id,
lsm_tree->merge_progressing));
(void)WT_ATOMIC_SUB4(chunk->refcnt, 1);
flushing = ref = 0;
compacting = 1;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
progress = lsm_tree->merge_progressing;
} else {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Compact flush retry %s chunk %u",
name, chunk->id));
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_FLUSH, WT_LSM_WORK_FORCE,
lsm_tree));
}
}
/*
* The compacting flag is cleared when no merges can be done.
* Ensure that we push through some aggressive merges before
* stopping otherwise we might not do merges that would
* span chunks with different generations.
*/
if (compacting && !F_ISSET(lsm_tree, WT_LSM_TREE_COMPACTING)) {
if (lsm_tree->merge_aggressiveness < 10 ||
(progress < lsm_tree->merge_progressing) ||
lsm_tree->merge_syncing) {
progress = lsm_tree->merge_progressing;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
lsm_tree->merge_aggressiveness = 10;
} else
break;
}
__wt_sleep(1, 0);
WT_ERR(__wt_seconds(session, &end));
if (session->compact->max_time > 0 &&
session->compact->max_time < (uint64_t)(end - begin)) {
WT_ERR(ETIMEDOUT);
}
/*
* Push merge operations while they are still getting work
* done. If we are pushing merges, make sure they are
* aggressive, to avoid duplicating effort.
*/
if (compacting)
#define COMPACT_PARALLEL_MERGES 5
for (i = lsm_tree->queue_ref;
i < COMPACT_PARALLEL_MERGES; i++) {
lsm_tree->merge_aggressiveness = 10;
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
}
}
err:
/* Ensure anything we set is cleared. */
if (ref)
(void)WT_ATOMIC_SUB4(chunk->refcnt, 1);
if (compacting) {
F_CLR(lsm_tree, WT_LSM_TREE_COMPACTING);
lsm_tree->merge_aggressiveness = 0;
}
if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Compact %s complete, return %d", name, ret));
__wt_lsm_tree_release(session, lsm_tree);
return (ret);
}
/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_DECL_RET;
WT_PAGE *page;
u_int sleep_cnt, wait_cnt;
int busy, force_attempts, oldgen;
for (force_attempts = oldgen = 0, wait_cnt = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, attempt to read it.
* Make sure there is space in the cache.
*/
WT_RET(__wt_cache_full_check(session));
WT_RET(__wt_cache_read(session, ref));
oldgen = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
break;
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory: get a hazard pointer, update
* the page's LRU and return. The expected reason we
* can't get a hazard pointer is because the page is
* being evicted; yield and try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy) {
WT_STAT_FAST_CONN_INCR(
session, page_busy_blocked);
break;
}
page = ref->page;
WT_ASSERT(session, page != NULL);
/*
* Forcibly evict pages that are too big.
*/
if (force_attempts < 10 &&
__evict_force_check(session, page, flags)) {
++force_attempts;
ret = __wt_page_release_evict(session, ref);
/* If forced eviction fails, stall. */
if (ret == EBUSY) {
ret = 0;
wait_cnt += 1000;
WT_STAT_FAST_CONN_INCR(session,
page_forcible_evict_blocked);
break;
} else
WT_RET(ret);
/*
* The result of a successful forced eviction
* is a page-state transition (potentially to
* an in-memory page we can use, or a restart
* return for our caller), continue the outer
* page-acquisition loop.
*/
continue;
}
/* Check if we need an autocommit transaction. */
if ((ret = __wt_txn_autocommit_check(session)) != 0) {
WT_TRET(__wt_hazard_clear(session, page));
return (ret);
}
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*
* Otherwise, update the page's read generation.
*/
if (oldgen && page->read_gen == WT_READGEN_NOTSET)
__wt_page_evict_soon(page);
else if (!LF_ISSET(WT_READ_NO_GEN) &&
page->read_gen != WT_READGEN_OLDEST &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_set(session);
return (0);
WT_ILLEGAL_VALUE(session);
}
/*
* We failed to get the page -- yield before retrying, and if
* we've yielded enough times, start sleeping so we don't burn
* CPU to no purpose.
*/
if (++wait_cnt < 1000)
__wt_yield();
else {
sleep_cnt = WT_MIN(wait_cnt, 10000);
wait_cnt *= 2;
WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
__wt_sleep(0, sleep_cnt);
}
}
}
/*
* Regularly create, open a cursor and drop a table.
* Measure how long each step takes, and flag an error if it exceeds the
* configured maximum.
*/
static void *
cycle_idle_tables(void *arg)
{
struct timespec start, stop;
CONFIG *cfg;
WT_SESSION *session;
WT_CURSOR *cursor;
int cycle_count, ret;
char uri[512];
cfg = (CONFIG *)arg;
cycle_count = 0;
if ((ret = cfg->conn->open_session(
cfg->conn, NULL, cfg->sess_config, &session)) != 0) {
lprintf(cfg, ret, 0,
"Error opening a session on %s", cfg->home);
return (NULL);
}
for (cycle_count = 0; cfg->idle_cycle_run; ++cycle_count) {
snprintf(uri, 512, "%s_cycle%07d", cfg->uris[0], cycle_count);
/* Don't busy cycle in this loop. */
__wt_sleep(1, 0);
/* Setup a start timer. */
if ((ret = __wt_epoch(NULL, &start)) != 0) {
lprintf(cfg, ret, 0,
"Get time failed in cycle_idle_tables.");
cfg->error = ret;
return (NULL);
}
/* Create a table. */
if ((ret = session->create(
session, uri, cfg->table_config)) != 0) {
if (ret == EBUSY)
continue;
lprintf(cfg, ret, 0,
"Table create failed in cycle_idle_tables.");
cfg->error = ret;
return (NULL);
}
if (check_timing(cfg, "create", start, &stop) != 0)
return (NULL);
start = stop;
/* Open and close cursor. */
if ((ret = session->open_cursor(
session, uri, NULL, NULL, &cursor)) != 0) {
lprintf(cfg, ret, 0,
"Cursor open failed in cycle_idle_tables.");
cfg->error = ret;
return (NULL);
}
if ((ret = cursor->close(cursor)) != 0) {
lprintf(cfg, ret, 0,
"Cursor close failed in cycle_idle_tables.");
cfg->error = ret;
return (NULL);
}
if (check_timing(cfg, "cursor", start, &stop) != 0)
return (NULL);
start = stop;
/*
* Drop the table. Keep retrying on EBUSY failure - it is an
* expected return when checkpoints are happening.
*/
while ((ret = session->drop(
session, uri, "force,checkpoint_wait=false")) == EBUSY)
__wt_sleep(1, 0);
if (ret != 0 && ret != EBUSY) {
lprintf(cfg, ret, 0,
"Table drop failed in cycle_idle_tables.");
cfg->error = ret;
return (NULL);
}
if (check_timing(cfg, "drop", start, &stop) != 0)
return (NULL);
}
return (NULL);
}
/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
u_int sleep_cnt, wait_cnt;
bool busy, cache_work, evict_soon, stalled;
int force_attempts;
btree = S2BT(session);
/*
* Ignore reads of pages already known to be in cache, otherwise the
* eviction server can dominate these statistics.
*/
if (!LF_ISSET(WT_READ_CACHE)) {
WT_STAT_FAST_CONN_INCR(session, cache_pages_requested);
WT_STAT_FAST_DATA_INCR(session, cache_pages_requested);
}
for (evict_soon = stalled = false,
force_attempts = 0, sleep_cnt = wait_cnt = 0;;) {
switch (ref->state) {
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_NO_EMPTY) &&
__wt_delete_page_skip(session, ref, false))
return (WT_NOTFOUND);
/* FALLTHROUGH */
case WT_REF_DISK:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, read it. If this thread is
* allowed to do eviction work, check for space in the
* cache.
*/
if (!LF_ISSET(WT_READ_NO_EVICT))
WT_RET(__wt_cache_eviction_check(
session, 1, NULL));
WT_RET(__page_read(session, ref));
/*
* If configured to not trash the cache, leave the page
* generation unset, we'll set it before returning to
* the oldest read generation, so the page is forcibly
* evicted as soon as possible. We don't do that set
* here because we don't want to evict the page before
* we "acquire" it.
*/
evict_soon = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on another thread's read, stall. */
WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
stalled = true;
break;
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on eviction, stall. */
WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
stalled = true;
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory.
*
* Get a hazard pointer if one is required. We cannot
* be evicting if no hazard pointer is required, we're
* done.
*/
if (F_ISSET(btree, WT_BTREE_IN_MEMORY))
goto skip_evict;
/*
* The expected reason we can't get a hazard pointer is
* because the page is being evicted, yield, try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy) {
WT_STAT_FAST_CONN_INCR(
session, page_busy_blocked);
break;
}
/*
* If eviction is configured for this file, check to see
* if the page qualifies for forced eviction and update
* the page's generation number. If eviction isn't being
* done on this file, we're done.
* In-memory split of large pages is allowed while
* no_eviction is set on btree, whereas reconciliation
* is not allowed.
*/
if (LF_ISSET(WT_READ_NO_EVICT) ||
F_ISSET(session, WT_SESSION_NO_EVICTION) ||
(F_ISSET(btree, WT_BTREE_NO_EVICTION) &&
!F_ISSET(btree, WT_BTREE_NO_RECONCILE)))
goto skip_evict;
/*
* Forcibly evict pages that are too big.
*/
if (force_attempts < 10 &&
__evict_force_check(session, ref)) {
++force_attempts;
ret = __wt_page_release_evict(session, ref);
/* If forced eviction fails, stall. */
if (ret == EBUSY) {
ret = 0;
WT_STAT_FAST_CONN_INCR(session,
page_forcible_evict_blocked);
stalled = true;
break;
}
WT_RET(ret);
/*
* The result of a successful forced eviction
* is a page-state transition (potentially to
* an in-memory page we can use, or a restart
* return for our caller), continue the outer
* page-acquisition loop.
*/
continue;
}
/*
* If we read the page and are configured to not trash
* the cache, and no other thread has already used the
* page, set the oldest read generation so the page is
* forcibly evicted as soon as possible.
*
* Otherwise, if we read the page, or, if configured to
* update the page's read generation and the page isn't
* already flagged for forced eviction, update the page
* read generation.
*/
page = ref->page;
if (page->read_gen == WT_READGEN_NOTSET) {
if (evict_soon)
__wt_page_evict_soon(session, ref);
else
__wt_cache_read_gen_new(session, page);
} else if (!LF_ISSET(WT_READ_NO_GEN))
__wt_cache_read_gen_bump(session, page);
skip_evict:
/*
* Check if we need an autocommit transaction.
* Starting a transaction can trigger eviction, so skip
* it if eviction isn't permitted.
*/
return (LF_ISSET(WT_READ_NO_EVICT) ? 0 :
__wt_txn_autocommit_check(session));
WT_ILLEGAL_VALUE(session);
}
/*
* We failed to get the page -- yield before retrying, and if
* we've yielded enough times, start sleeping so we don't burn
* CPU to no purpose.
*/
if (stalled)
wait_cnt += WT_THOUSAND;
else if (++wait_cnt < WT_THOUSAND) {
__wt_yield();
continue;
}
/*
* If stalling and this thread is allowed to do eviction work,
* check if the cache needs help. If we do work for the cache,
* substitute that for a sleep.
*/
if (!LF_ISSET(WT_READ_NO_EVICT)) {
WT_RET(
__wt_cache_eviction_check(session, 1, &cache_work));
if (cache_work)
continue;
}
sleep_cnt = WT_MIN(sleep_cnt + WT_THOUSAND, 10000);
WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
__wt_sleep(0, sleep_cnt);
}
}
/*
* __wt_log_slot_join --
* Join a consolidated logging slot.
*/
void
__wt_log_slot_join(WT_SESSION_IMPL *session, uint64_t mysize,
uint32_t flags, WT_MYSLOT *myslot)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_LOGSLOT *slot;
uint64_t time_start, time_stop, usecs;
int64_t flag_state, new_state, old_state, released;
int32_t join_offset, new_join, wait_cnt;
bool closed, diag_yield, raced, slept, unbuffered, yielded;
conn = S2C(session);
log = conn->log;
time_start = time_stop = 0;
WT_ASSERT(session, !F_ISSET(session, WT_SESSION_LOCKED_SLOT));
WT_ASSERT(session, mysize != 0);
/*
* There should almost always be a slot open.
*/
unbuffered = yielded = false;
closed = raced = slept = false;
wait_cnt = 0;
#ifdef HAVE_DIAGNOSTIC
diag_yield = (++log->write_calls % 7) == 0;
if ((log->write_calls % WT_THOUSAND) == 0 ||
mysize > WT_LOG_SLOT_BUF_MAX) {
#else
diag_yield = false;
if (mysize > WT_LOG_SLOT_BUF_MAX) {
#endif
unbuffered = true;
F_SET(myslot, WT_MYSLOT_UNBUFFERED);
}
for (;;) {
WT_BARRIER();
slot = log->active_slot;
old_state = slot->slot_state;
if (WT_LOG_SLOT_OPEN(old_state)) {
/*
* Try to join our size into the existing size and
* atomically write it back into the state.
*/
flag_state = WT_LOG_SLOT_FLAGS(old_state);
released = WT_LOG_SLOT_RELEASED(old_state);
join_offset = WT_LOG_SLOT_JOINED(old_state);
if (unbuffered)
new_join = join_offset + WT_LOG_SLOT_UNBUFFERED;
else
new_join = join_offset + (int32_t)mysize;
new_state = (int64_t)WT_LOG_SLOT_JOIN_REL(
(int64_t)new_join, (int64_t)released,
(int64_t)flag_state);
/*
* Braces used due to potential empty body warning.
*/
if (diag_yield) {
WT_DIAGNOSTIC_YIELD;
}
/*
* Attempt to swap our size into the state.
*/
if (__wt_atomic_casiv64(
&slot->slot_state, old_state, new_state))
break;
WT_STAT_CONN_INCR(session, log_slot_races);
raced = true;
} else {
WT_STAT_CONN_INCR(session, log_slot_active_closed);
closed = true;
++wait_cnt;
}
if (!yielded)
time_start = __wt_clock(session);
yielded = true;
/*
* The slot is no longer open or we lost the race to
* update it. Yield and try again.
*/
if (wait_cnt < WT_THOUSAND)
__wt_yield();
else {
__wt_sleep(0, WT_THOUSAND);
slept = true;
}
}
/*
* We joined this slot. Fill in our information to return to
* the caller.
*/
if (!yielded)
WT_STAT_CONN_INCR(session, log_slot_immediate);
else {
WT_STAT_CONN_INCR(session, log_slot_yield);
time_stop = __wt_clock(session);
usecs = WT_CLOCKDIFF_US(time_stop, time_start);
WT_STAT_CONN_INCRV(session, log_slot_yield_duration, usecs);
if (closed)
WT_STAT_CONN_INCR(session, log_slot_yield_close);
if (raced)
WT_STAT_CONN_INCR(session, log_slot_yield_race);
if (slept)
WT_STAT_CONN_INCR(session, log_slot_yield_sleep);
}
if (LF_ISSET(WT_LOG_DSYNC | WT_LOG_FSYNC))
F_SET(slot, WT_SLOT_SYNC_DIR);
if (LF_ISSET(WT_LOG_FLUSH))
F_SET(slot, WT_SLOT_FLUSH);
if (LF_ISSET(WT_LOG_FSYNC))
F_SET(slot, WT_SLOT_SYNC);
if (F_ISSET(myslot, WT_MYSLOT_UNBUFFERED)) {
WT_ASSERT(session, slot->slot_unbuffered == 0);
WT_STAT_CONN_INCR(session, log_slot_unbuffered);
slot->slot_unbuffered = (int64_t)mysize;
}
myslot->slot = slot;
myslot->offset = join_offset;
myslot->end_offset = (wt_off_t)((uint64_t)join_offset + mysize);
}
/*
* __wt_log_slot_release --
* Each thread in a consolidated group releases its portion to
* signal it has completed copying its piece of the log into
* the memory buffer.
*/
int64_t
__wt_log_slot_release(WT_MYSLOT *myslot, int64_t size)
{
WT_LOGSLOT *slot;
wt_off_t cur_offset, my_start;
int64_t my_size, rel_size;
slot = myslot->slot;
my_start = slot->slot_start_offset + myslot->offset;
/*
* We maintain the last starting offset within this slot.
* This is used to know the offset of the last record that
* was written rather than the beginning record of the slot.
*/
while ((cur_offset = slot->slot_last_offset) < my_start) {
/*
* Set our offset if we are larger.
*/
if (__wt_atomic_casiv64(
&slot->slot_last_offset, cur_offset, my_start))
break;
/*
* If we raced another thread updating this, try again.
*/
WT_BARRIER();
}
/*
* Add my size into the state and return the new size.
*/
rel_size = size;
if (F_ISSET(myslot, WT_MYSLOT_UNBUFFERED))
rel_size = WT_LOG_SLOT_UNBUFFERED;
my_size = (int64_t)WT_LOG_SLOT_JOIN_REL((int64_t)0, rel_size, 0);
return (__wt_atomic_addiv64(&slot->slot_state, my_size));
}
/*
* __wt_page_out --
* Discard an in-memory page, freeing all memory associated with it.
*/
void
__wt_page_out(WT_SESSION_IMPL *session, WT_PAGE **pagep)
{
WT_PAGE *page;
WT_PAGE_HEADER *dsk;
WT_PAGE_MODIFY *mod;
/*
* Kill our caller's reference, do our best to catch races.
*/
page = *pagep;
*pagep = NULL;
if (F_ISSET(session->dhandle, WT_DHANDLE_DEAD))
__wt_page_modify_clear(session, page);
/*
* We should never discard:
* - a dirty page,
* - a page queued for eviction, or
* - a locked page.
*/
WT_ASSERT(session, !__wt_page_is_modified(page));
WT_ASSERT(session, !F_ISSET_ATOMIC(page, WT_PAGE_EVICT_LRU));
WT_ASSERT(session, !__wt_fair_islocked(session, &page->page_lock));
#ifdef HAVE_DIAGNOSTIC
{
WT_HAZARD *hp;
int i;
/*
* Make sure no other thread has a hazard pointer on the page we are
* about to discard. This is complicated by the fact that readers
* publish their hazard pointer before re-checking the page state, so
* our check can race with readers without indicating a real problem.
* Wait for up to a second for hazard pointers to be cleared.
*/
for (hp = NULL, i = 0; i < 100; i++) {
if ((hp = __wt_page_hazard_check(session, page)) == NULL)
break;
__wt_sleep(0, 10000);
}
if (hp != NULL)
__wt_errx(session,
"discarded page has hazard pointer: (%p: %s, line %d)",
hp->page, hp->file, hp->line);
WT_ASSERT(session, hp == NULL);
}
#endif
/*
* If a root page split, there may be one or more pages linked from the
* page; walk the list, discarding pages.
*/
switch (page->type) {
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
mod = page->modify;
if (mod != NULL && mod->mod_root_split != NULL)
__wt_page_out(session, &mod->mod_root_split);
break;
}
/* Update the cache's information. */
__wt_cache_page_evict(session, page);
/*
* If discarding the page as part of process exit, the application may
* configure to leak the memory rather than do the work.
*/
if (F_ISSET(S2C(session), WT_CONN_LEAK_MEMORY))
return;
/* Free the page modification information. */
if (page->modify != NULL)
__free_page_modify(session, page);
switch (page->type) {
case WT_PAGE_COL_FIX:
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
__free_page_int(session, page);
break;
case WT_PAGE_COL_VAR:
__free_page_col_var(session, page);
break;
case WT_PAGE_ROW_LEAF:
__free_page_row_leaf(session, page);
break;
}
/* Discard any disk image. */
dsk = (WT_PAGE_HEADER *)page->dsk;
if (F_ISSET_ATOMIC(page, WT_PAGE_DISK_ALLOC))
__wt_overwrite_and_free_len(session, dsk, dsk->mem_size);
if (F_ISSET_ATOMIC(page, WT_PAGE_DISK_MAPPED))
(void)__wt_mmap_discard(session, dsk, dsk->mem_size);
__wt_overwrite_and_free(session, page);
}
/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
u_int sleep_cnt, wait_cnt;
int busy, cache_work, force_attempts, oldgen, stalled;
btree = S2BT(session);
stalled = 0;
for (force_attempts = oldgen = 0, sleep_cnt = wait_cnt = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, read it. If this thread is
* allowed to do eviction work, check for space in the
* cache.
*/
if (!LF_ISSET(WT_READ_NO_EVICT))
WT_RET(__wt_cache_eviction_check(
session, 1, NULL));
WT_RET(__page_read(session, ref));
oldgen = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on another thread's read, stall. */
WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
stalled = 1;
break;
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on eviction, stall. */
WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
stalled = 1;
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory.
*
* Get a hazard pointer if one is required. We cannot
* be evicting if no hazard pointer is required, we're
* done.
*/
if (F_ISSET(btree, WT_BTREE_IN_MEMORY))
goto skip_evict;
/*
* The expected reason we can't get a hazard pointer is
* because the page is being evicted, yield, try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy) {
WT_STAT_FAST_CONN_INCR(
session, page_busy_blocked);
break;
}
/*
* If eviction is configured for this file, check to see
* if the page qualifies for forced eviction and update
* the page's generation number. If eviction isn't being
* done on this file, we're done.
*/
if (LF_ISSET(WT_READ_NO_EVICT) ||
F_ISSET(session, WT_SESSION_NO_EVICTION) ||
F_ISSET(btree, WT_BTREE_NO_EVICTION))
goto skip_evict;
/*
* Forcibly evict pages that are too big.
*/
page = ref->page;
if (force_attempts < 10 &&
__evict_force_check(session, page)) {
++force_attempts;
ret = __wt_page_release_evict(session, ref);
/* If forced eviction fails, stall. */
if (ret == EBUSY) {
ret = 0;
WT_STAT_FAST_CONN_INCR(session,
page_forcible_evict_blocked);
stalled = 1;
break;
}
WT_RET(ret);
/*
* The result of a successful forced eviction
* is a page-state transition (potentially to
* an in-memory page we can use, or a restart
* return for our caller), continue the outer
* page-acquisition loop.
*/
continue;
}
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*
* Otherwise, update the page's read generation.
*/
if (oldgen && page->read_gen == WT_READGEN_NOTSET)
__wt_page_evict_soon(page);
else if (!LF_ISSET(WT_READ_NO_GEN) &&
page->read_gen != WT_READGEN_OLDEST &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_bump(session);
skip_evict:
/*
* Check if we need an autocommit transaction.
* Starting a transaction can trigger eviction, so skip
* it if eviction isn't permitted.
*/
return (LF_ISSET(WT_READ_NO_EVICT) ? 0 :
__wt_txn_autocommit_check(session));
WT_ILLEGAL_VALUE(session);
}
/*
* We failed to get the page -- yield before retrying, and if
* we've yielded enough times, start sleeping so we don't burn
* CPU to no purpose.
*/
if (stalled)
wait_cnt += 1000;
else if (++wait_cnt < 1000) {
__wt_yield();
continue;
}
/*
* If stalling and this thread is allowed to do eviction work,
* check if the cache needs help. If we do work for the cache,
* substitute that for a sleep.
*/
if (!LF_ISSET(WT_READ_NO_EVICT)) {
WT_RET(
__wt_cache_eviction_check(session, 1, &cache_work));
if (cache_work)
continue;
}
sleep_cnt = WT_MIN(sleep_cnt + 1000, 10000);
WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
__wt_sleep(0, sleep_cnt);
}
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, int update, u_int start_chunk, uint32_t start_id)
{
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
int locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
lsm_tree = clsm->lsm_tree;
chunk = NULL;
ckpt_cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/* Copy the key, so we don't lose the cursor position. */
if (F_ISSET(c, WT_CURSTD_KEY_INT) && !WT_DATA_IN_ITEM(&c->key))
WT_RET(__wt_buf_set(
session, &c->key, c->key.data, c->key.size));
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
if (update) {
if (txn->isolation == TXN_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
} else
F_SET(clsm, WT_CLSM_OPEN_READ);
WT_RET(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
/*
* If there is no in-memory chunk in the tree for an update operation,
* create one.
*
* !!!
* It is exceeding unlikely that we get here at all, but if we were to
* switch chunks in this thread and our transaction roll back, it would
* leave the metadata inconsistent. Signal for the LSM worker thread
* to create the chunk instead to avoid the issue.
*/
if (update && (lsm_tree->nchunks == 0 ||
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) == NULL ||
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))) {
/* Release our lock because switch will get a write lock. */
F_SET(lsm_tree, WT_LSM_TREE_NEED_SWITCH);
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__wt_cond_signal(session, lsm_tree->work_cond));
/*
* Give the worker thread a chance to run before locking the
* tree again -- we will loop in __clsm_enter until there is an
* in-memory chunk in the tree.
*/
__wt_sleep(0, 1000);
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
}
F_SET(session, WT_SESSION_NO_CACHE_CHECK);
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id) {
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->txnid_max));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->txnid_max[ngood - 1] =
chunk->txnid_max;
if (__wt_txn_visible_all(
session, chunk->txnid_max))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need. If the cursor is a
* pure update cursor, close everything -- we usually only need
* a single chunk open in that case and we haven't walked all
* of the other slots in the loop above.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0)
ngood = 0;
if (clsm->cursors != NULL && ngood < clsm->nchunks) {
saved_gen = lsm_tree->dsk_gen;
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__clsm_close_cursors(
clsm, ngood, clsm->nchunks));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->txnid_max[i] = chunk->txnid_max;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL && !F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
WT_WITH_BTREE(session, ((WT_CURSOR_BTREE *)(primary))->btree,
__wt_btree_evictable(session, 0));
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err: F_CLR(session, WT_SESSION_NO_CACHE_CHECK);
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
return (ret);
}
