/*
* __wt_lsm_work_switch --
* Do a switch if the LSM tree needs one.
*/
int
__wt_lsm_work_switch(
WT_SESSION_IMPL *session, WT_LSM_WORK_UNIT **entryp, bool *ran)
{
WT_DECL_RET;
WT_LSM_WORK_UNIT *entry;
/* We've become responsible for freeing the work unit. */
entry = *entryp;
*ran = false;
*entryp = NULL;
if (F_ISSET(entry->lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_lsm_tree_switch(session, entry->lsm_tree));
/* Failing to complete the switch is fine */
if (ret == EBUSY) {
if (F_ISSET(entry->lsm_tree, WT_LSM_TREE_NEED_SWITCH))
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_SWITCH, 0, entry->lsm_tree));
ret = 0;
} else
*ran = true;
}
err: __wt_lsm_manager_free_work_unit(session, entry);
return (ret);
}
/*
* __wt_clsm_open_bulk --
* WT_SESSION->open_cursor method for LSM bulk cursors.
*/
int
__wt_clsm_open_bulk(WT_CURSOR_LSM *clsm, const char *cfg[])
{
WT_CURSOR *cursor, *bulk_cursor;
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
bulk_cursor = NULL;
cursor = &clsm->iface;
lsm_tree = clsm->lsm_tree;
session = (WT_SESSION_IMPL *)clsm->iface.session;
F_SET(clsm, WT_CLSM_BULK);
/* Bulk cursors are limited to insert and close. */
__wt_cursor_set_notsup(cursor);
cursor->insert = __clsm_insert_bulk;
cursor->close = __clsm_close_bulk;
/*
* Setup the first chunk in the tree. This is the only time we switch
* without using the LSM worker threads, it's safe to do here since
* we have an exclusive lock on the LSM tree. We need to do this
* switch inline, since switch needs a schema lock and online index
* creation opens a bulk cursor while holding the schema lock.
*/
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_lsm_tree_switch(session, lsm_tree));
WT_RET(ret);
/*
* Open a bulk cursor on the first chunk, it's not a regular LSM chunk
* cursor, but use the standard storage locations. Allocate the space
* for a bloom filter - it makes cleanup simpler. Cleaned up by
* cursor close on error.
*/
WT_RET(__wt_calloc_one(session, &clsm->blooms));
clsm->bloom_alloc = 1;
WT_RET(__wt_calloc_one(session, &clsm->cursors));
clsm->cursor_alloc = 1;
clsm->nchunks = 1;
/*
* Open a bulk cursor on the first chunk in the tree - take a read
* lock on the LSM tree while we are opening the chunk, to ensure
* that the first chunk has been fully created before we succeed.
* Pass through the application config to ensure the tree is open
* for bulk access.
*/
WT_RET(__wt_open_cursor(session,
lsm_tree->chunk[0]->uri, &clsm->iface, cfg, &bulk_cursor));
clsm->cursors[0] = bulk_cursor;
/* LSM cursors are always raw */
F_SET(bulk_cursor, WT_CURSTD_RAW);
return (0);
}
/*
* __lsm_tree_open --
* Open an LSM tree structure.
*/
static int
__lsm_tree_open(
WT_SESSION_IMPL *session, const char *uri, WT_LSM_TREE **treep)
{
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_SCHEMA_LOCKED));
/* Make sure no one beat us to it. */
TAILQ_FOREACH(lsm_tree, &S2C(session)->lsmqh, q)
if (strcmp(uri, lsm_tree->name) == 0) {
*treep = lsm_tree;
return (0);
}
/* Try to open the tree. */
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__wt_rwlock_alloc(session, "lsm tree", &lsm_tree->rwlock));
WT_ERR(__wt_cond_alloc(session, "lsm ckpt", 0, &lsm_tree->work_cond));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_lsm_meta_read(session, lsm_tree));
/*
* Sanity check the configuration. Do it now since this is the first
* time we have the LSM tree configuration.
*/
WT_ERR(__lsm_tree_open_check(session, lsm_tree));
if (lsm_tree->nchunks == 0) {
F_SET(lsm_tree, WT_LSM_TREE_NEED_SWITCH);
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
}
/* Set the generation number so cursors are opened on first usage. */
lsm_tree->dsk_gen = 1;
/* Now the tree is setup, make it visible to others. */
lsm_tree->refcnt = 1;
TAILQ_INSERT_HEAD(&S2C(session)->lsmqh, lsm_tree, q);
F_SET(lsm_tree, WT_LSM_TREE_OPEN);
WT_ERR(__lsm_tree_start_worker(session, lsm_tree));
*treep = lsm_tree;
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
return (ret);
}
/*
* __lsm_tree_open --
* Open an LSM tree structure.
*/
static int
__lsm_tree_open(
WT_SESSION_IMPL *session, const char *uri, WT_LSM_TREE **treep)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
conn = S2C(session);
lsm_tree = NULL;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_SCHEMA_LOCKED));
/* Start the LSM manager thread if it isn't running. */
if (WT_ATOMIC_CAS4(conn->lsm_manager.lsm_workers, 0, 1))
WT_RET(__wt_lsm_manager_start(session));
/* Make sure no one beat us to it. */
TAILQ_FOREACH(lsm_tree, &S2C(session)->lsmqh, q)
if (strcmp(uri, lsm_tree->name) == 0) {
*treep = lsm_tree;
return (0);
}
/* Try to open the tree. */
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__wt_rwlock_alloc(session, &lsm_tree->rwlock, "lsm tree"));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_lsm_meta_read(session, lsm_tree));
/*
* Sanity check the configuration. Do it now since this is the first
* time we have the LSM tree configuration.
*/
WT_ERR(__lsm_tree_open_check(session, lsm_tree));
if (lsm_tree->nchunks == 0) {
F_SET(lsm_tree, WT_LSM_TREE_NEED_SWITCH);
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
}
/* Set the generation number so cursors are opened on first usage. */
lsm_tree->dsk_gen = 1;
/*
* Setup reference counting. Use separate reference counts for tree
* handles and queue entries, so that queue entries don't interfere
* with getting handles exclusive.
*/
lsm_tree->refcnt = 1;
lsm_tree->queue_ref = 0;
/* Set a flush timestamp as a baseline. */
WT_ERR(__wt_epoch(session, &lsm_tree->last_flush_ts));
/* Now the tree is setup, make it visible to others. */
TAILQ_INSERT_HEAD(&S2C(session)->lsmqh, lsm_tree, q);
F_SET(lsm_tree, WT_LSM_TREE_ACTIVE | WT_LSM_TREE_OPEN);
*treep = lsm_tree;
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
return (ret);
}
/*
* __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 there is a
* transaction in progress and it rolls back, it would leave the
* metadata inconsistent.
*/
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. */
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
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.
*
* 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 (clsm->cursors != NULL && (ngood < clsm->nchunks ||
(!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0))) {
saved_gen = lsm_tree->dsk_gen;
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0)
WT_ERR(__clsm_close_cursors(
clsm, 0, nchunks - nupdates));
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);
}
/*
* __wt_lsm_merge_worker --
* The merge worker thread for an LSM tree, responsible for merging
* on-disk trees.
*/
void *
__wt_lsm_merge_worker(void *vargs)
{
WT_DECL_RET;
WT_LSM_WORKER_ARGS *args;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
u_int aggressive, chunk_wait, id, old_aggressive, stallms;
int progress;
args = vargs;
lsm_tree = args->lsm_tree;
id = args->id;
session = lsm_tree->worker_sessions[id];
__wt_free(session, args);
aggressive = stallms = 0;
while (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING)) {
/*
* Help out with switching chunks in case the checkpoint worker
* is busy.
*/
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret =
__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(ret);
}
progress = 0;
/* Clear any state from previous worker thread iterations. */
session->dhandle = NULL;
/* Try to create a Bloom filter. */
if (__lsm_bloom_work(session, lsm_tree) == 0)
progress = 1;
/* If we didn't create a Bloom filter, try to merge. */
if (progress == 0 &&
__wt_lsm_merge(session, lsm_tree, id, aggressive) == 0)
progress = 1;
/* Clear any state from previous worker thread iterations. */
WT_CLEAR_BTREE_IN_SESSION(session);
/*
* Only have one thread freeing old chunks, and only if there
* are chunks to free.
*/
if (id == 0 && lsm_tree->nold_chunks > 0 &&
__lsm_free_chunks(session, lsm_tree) == 0)
progress = 1;
if (progress)
stallms = 0;
else if (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING) &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
/* Poll 10 times per second. */
WT_ERR_TIMEDOUT_OK(__wt_cond_wait(
session, lsm_tree->work_cond, 100000));
stallms += 100;
/*
* Get aggressive if more than enough chunks for a
* merge should have been created while we waited.
* Use 10 seconds as a default if we don't have an
* estimate.
*/
chunk_wait = stallms / (lsm_tree->chunk_fill_ms == 0 ?
10000 : lsm_tree->chunk_fill_ms);
old_aggressive = aggressive;
aggressive = chunk_wait / lsm_tree->merge_min;
if (aggressive > old_aggressive)
WT_VERBOSE_ERR(session, lsm,
"LSM merge got aggressive (%u), "
"%u / %" PRIu64,
aggressive, stallms,
lsm_tree->chunk_fill_ms);
}
}
if (0) {
err: __wt_err(session, ret, "LSM merge worker failed");
}
return (NULL);
}
/*
* __wt_lsm_checkpoint_worker --
* A worker thread for an LSM tree, responsible for flushing new chunks to
* disk.
*/
void *
__wt_lsm_checkpoint_worker(void *arg)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_LSM_WORKER_COOKIE cookie;
WT_SESSION_IMPL *session;
WT_TXN_ISOLATION saved_isolation;
u_int i, j;
int locked;
lsm_tree = arg;
session = lsm_tree->ckpt_session;
WT_CLEAR(cookie);
while (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING)) {
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret =
__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(ret);
}
WT_ERR(__lsm_copy_chunks(session, lsm_tree, &cookie, 0));
/* Write checkpoints in all completed files. */
for (i = 0, j = 0; i < cookie.nchunks - 1; i++) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_WORKING))
goto err;
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
break;
chunk = cookie.chunk_array[i];
/* Stop if a running transaction needs the chunk. */
__wt_txn_update_oldest(session);
if (!__wt_txn_visible_all(session, chunk->txnid_max))
break;
/*
* If the chunk is already checkpointed, make sure it
* is also evicted. Either way, there is no point
* trying to checkpoint it again.
*/
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_ONDISK)) {
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_EVICTED))
continue;
if ((ret = __lsm_discard_handle(
session, chunk->uri, NULL)) == 0)
F_SET_ATOMIC(
chunk, WT_LSM_CHUNK_EVICTED);
else if (ret == EBUSY)
ret = 0;
else
WT_ERR_MSG(session, ret,
"discard handle");
continue;
}
WT_VERBOSE_ERR(session, lsm,
"LSM worker flushing %u", i);
/*
* Flush the file before checkpointing: this is the
* expensive part in terms of I/O: do it without
* holding the schema lock.
*
* Use the special eviction isolation level to avoid
* interfering with an application checkpoint: we have
* already checked that all of the updates in this
* chunk are globally visible.
*
* !!! We can wait here for checkpoints and fsyncs to
* complete, which can be a long time.
*
* Don't keep waiting for the lock if application
* threads are waiting for a switch. Don't skip
* flushing the leaves either: that just means we'll
* hold the schema lock for (much) longer, which blocks
* the world.
*/
WT_ERR(__wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0));
for (locked = 0;
!locked && ret == 0 &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH);) {
if ((ret = __wt_spin_trylock(session,
&S2C(session)->checkpoint_lock)) == 0)
locked = 1;
else if (ret == EBUSY) {
__wt_yield();
ret = 0;
}
}
if (locked) {
saved_isolation = session->txn.isolation;
session->txn.isolation = TXN_ISO_EVICTION;
ret = __wt_bt_cache_op(
session, NULL, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
__wt_spin_unlock(
session, &S2C(session)->checkpoint_lock);
}
WT_TRET(__wt_session_release_btree(session));
WT_ERR(ret);
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
break;
WT_VERBOSE_ERR(session, lsm,
"LSM worker checkpointing %u", i);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, chunk->uri,
__wt_checkpoint, NULL, NULL, 0));
if (ret != 0) {
__wt_err(session, ret, "LSM checkpoint");
break;
}
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
/*
* Clear the "cache resident" flag so the primary can
* be evicted and eventually closed. Only do this once
* the checkpoint has succeeded: otherwise, accessing
* the leaf page during the checkpoint can trigger
* forced eviction.
*/
WT_ERR(__wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0));
__wt_btree_evictable(session, 1);
WT_ERR(__wt_session_release_btree(session));
++j;
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
F_SET_ATOMIC(chunk, WT_LSM_CHUNK_ONDISK);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
/* Update the throttle time. */
__wt_lsm_tree_throttle(session, lsm_tree);
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
if (ret != 0) {
__wt_err(session, ret,
"LSM checkpoint metadata write");
break;
}
WT_VERBOSE_ERR(session, lsm,
"LSM worker checkpointed %u", i);
}
__lsm_unpin_chunks(session, &cookie);
if (j == 0 && F_ISSET(lsm_tree, WT_LSM_TREE_WORKING) &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
WT_ERR_TIMEDOUT_OK(__wt_cond_wait(
session, lsm_tree->work_cond, 100000));
}
err: __lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
/*
* The thread will only exit with failure if we run out of memory or
* there is some other system driven failure. We can't keep going
* after such a failure - ensure WiredTiger shuts down.
*/
if (ret != 0 && ret != WT_NOTFOUND)
WT_PANIC_ERR(session, ret,
"Shutting down LSM checkpoint utility thread");
return (NULL);
}
