/*
* __wt_lsm_checkpoint_chunk --
* Flush a single LSM chunk to disk.
*/
int
__wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_DECL_RET;
WT_TXN_ISOLATION saved_isolation;
bool flush_set;
flush_set = false;
/*
* 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(chunk, WT_LSM_CHUNK_ONDISK) &&
!F_ISSET(chunk, WT_LSM_CHUNK_STABLE) &&
!chunk->evicted) {
WT_WITH_HANDLE_LIST_LOCK(session,
ret = __lsm_discard_handle(session, chunk->uri, NULL));
if (ret == 0)
chunk->evicted = 1;
else if (ret == EBUSY)
ret = 0;
else
WT_RET_MSG(session, ret, "discard handle");
}
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s already on disk",
chunk->uri);
return (0);
}
/* Stop if a running transaction needs the chunk. */
WT_RET(__wt_txn_update_oldest(
session, WT_TXN_OLDEST_STRICT | WT_TXN_OLDEST_WAIT));
if (chunk->switch_txn == WT_TXN_NONE ||
!__wt_txn_visible_all(session, chunk->switch_txn)) {
__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s: running transaction, return",
chunk->uri);
return (0);
}
if (!__wt_atomic_cas8(&chunk->flushing, 0, 1))
return (0);
flush_set = true;
__wt_verbose(session, WT_VERB_LSM, "LSM worker flushing %s",
chunk->uri);
/*
* Flush the file before checkpointing: this is the expensive part in
* terms of I/O.
*
* !!!
* We can wait here for checkpoints and fsyncs to complete, which can
* take a long time.
*/
if ((ret = __wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0)) == 0) {
/*
* Set read-uncommitted: we have already checked that all of the
* updates in this chunk are globally visible, use the cheapest
* possible check in reconciliation.
*/
saved_isolation = session->txn.isolation;
session->txn.isolation = WT_ISO_READ_UNCOMMITTED;
ret = __wt_cache_op(session, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
WT_TRET(__wt_session_release_btree(session));
}
WT_ERR(ret);
__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s",
chunk->uri);
/*
* Turn on metadata tracking to ensure the checkpoint gets the
* necessary handle locks.
*
* Ensure that we don't race with a running checkpoint: the checkpoint
* lock protects against us racing with an application checkpoint in
* this chunk. Don't wait for it, though: checkpoints can take a long
* time, and our checkpoint operation should be very quick.
*/
WT_ERR(__wt_meta_track_on(session));
WT_WITH_CHECKPOINT_LOCK(session, ret,
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_schema_worker(
session, chunk->uri, __wt_checkpoint, NULL, NULL, 0)));
WT_TRET(__wt_meta_track_off(session, false, ret != 0));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM checkpoint");
/* Now the file is written, get the chunk size. */
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
/* Update the flush timestamp to help track ongoing progress. */
WT_ERR(__wt_epoch(session, &lsm_tree->last_flush_ts));
++lsm_tree->chunks_flushed;
/* Lock the tree, mark the chunk as on disk and update the metadata. */
__wt_lsm_tree_writelock(session, lsm_tree);
F_SET(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, true);
__wt_lsm_tree_writeunlock(session, lsm_tree);
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM metadata write");
WT_PUBLISH(chunk->flushing, 0);
flush_set = false;
/*
* Clear the no-eviction 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, true);
WT_ERR(__wt_session_release_btree(session));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointed %s",
chunk->uri);
/* Schedule a bloom filter create for our newly flushed chunk. */
if (!FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF))
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_BLOOM, 0, lsm_tree));
else
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
err: if (flush_set)
WT_PUBLISH(chunk->flushing, 0);
return (ret);
}
/*
* __lsm_bloom_create --
* Create a bloom filter for a chunk of the LSM tree that has been
* checkpointed but not yet been merged.
*/
static int
__lsm_bloom_create(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk, u_int chunk_off)
{
WT_BLOOM *bloom;
WT_CURSOR *src;
WT_DECL_RET;
WT_ITEM key;
uint64_t insert_count;
WT_RET(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
bloom = NULL;
/*
* This is merge-like activity, and we don't want compacts to give up
* because we are creating a bunch of bloom filters before merging.
*/
++lsm_tree->merge_progressing;
WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config, chunk->count,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
/* Open a special merge cursor just on this chunk. */
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, chunk_off, chunk->id, 1));
/*
* Setup so that we don't hold pages we read into cache, and so
* that we don't get stuck if the cache is full. If we allow
* ourselves to get stuck creating bloom filters, the entire tree
* can stall since there may be no worker threads available to flush.
*/
F_SET(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
WT_ERR(src->get_key(src, &key));
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_TRET(src->close(src));
WT_TRET(__wt_bloom_finalize(bloom));
WT_ERR(ret);
F_CLR(session, WT_SESSION_NO_CACHE);
/* Load the new Bloom filter into cache. */
WT_CLEAR(key);
WT_ERR_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
__wt_verbose(session, WT_VERB_LSM,
"LSM worker created bloom filter %s. "
"Expected %" PRIu64 " items, got %" PRIu64,
chunk->bloom_uri, chunk->count, insert_count);
/* Ensure the bloom filter is in the metadata. */
__wt_lsm_tree_writelock(session, lsm_tree);
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
__wt_lsm_tree_writeunlock(session, lsm_tree);
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM bloom worker metadata write");
err: if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __wt_lsm_merge --
* Merge a set of chunks of an LSM tree.
*/
int
__wt_lsm_merge(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, u_int id)
{
WT_BLOOM *bloom;
WT_CURSOR *dest, *src;
WT_DECL_RET;
WT_ITEM key, value;
WT_LSM_CHUNK *chunk;
uint32_t generation;
uint64_t insert_count, record_count;
u_int dest_id, end_chunk, i, nchunks, start_chunk, start_id, verb;
int tret;
bool created_chunk, create_bloom, locked, in_sync;
const char *cfg[3];
const char *drop_cfg[] =
{ WT_CONFIG_BASE(session, WT_SESSION_drop), "force", NULL };
bloom = NULL;
chunk = NULL;
dest = src = NULL;
start_id = 0;
created_chunk = create_bloom = locked = in_sync = false;
/* Fast path if it's obvious no merges could be done. */
if (lsm_tree->nchunks < lsm_tree->merge_min &&
lsm_tree->merge_aggressiveness < WT_LSM_AGGRESSIVE_THRESHOLD)
return (WT_NOTFOUND);
/*
* Use the lsm_tree lock to read the chunks (so no switches occur), but
* avoid holding it while the merge is in progress: that may take a
* long time.
*/
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
WT_ERR(__lsm_merge_span(session,
lsm_tree, id, &start_chunk, &end_chunk, &record_count));
nchunks = (end_chunk + 1) - start_chunk;
WT_ASSERT(session, nchunks > 0);
start_id = lsm_tree->chunk[start_chunk]->id;
/* Find the merge generation. */
for (generation = 0, i = 0; i < nchunks; i++)
generation = WT_MAX(generation,
lsm_tree->chunk[start_chunk + i]->generation + 1);
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
locked = false;
/* Allocate an ID for the merge. */
dest_id = __wt_atomic_add32(&lsm_tree->last, 1);
/*
* We only want to do the chunk loop if we're running with verbose,
* so we wrap these statements in the conditional. Avoid the loop
* in the normal path.
*/
if (WT_VERBOSE_ISSET(session, WT_VERB_LSM)) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s chunks %u-%u into %u (%" PRIu64 " records)"
", generation %" PRIu32,
lsm_tree->name,
start_chunk, end_chunk, dest_id, record_count, generation));
for (verb = start_chunk; verb <= end_chunk; verb++)
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s: Chunk[%u] id %u, gen: %" PRIu32
", size: %" PRIu64 ", records: %" PRIu64,
lsm_tree->name, verb, lsm_tree->chunk[verb]->id,
lsm_tree->chunk[verb]->generation,
lsm_tree->chunk[verb]->size,
lsm_tree->chunk[verb]->count));
}
WT_ERR(__wt_calloc_one(session, &chunk));
created_chunk = true;
chunk->id = dest_id;
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_MERGED) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST) ||
start_chunk > 0) && record_count > 0)
create_bloom = true;
/*
* Special setup for the merge cursor:
* first, reset to open the dependent cursors;
* then restrict the cursor to a specific number of chunks;
* then set MERGE so the cursor doesn't track updates to the tree.
*/
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, start_chunk, start_id, nchunks));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
WT_ERR(ret);
if (create_bloom) {
WT_ERR(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
WT_ERR(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config,
record_count, lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count, &bloom));
}
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = "bulk,raw,skip_sort_check";
cfg[2] = NULL;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
#define LSM_MERGE_CHECK_INTERVAL WT_THOUSAND
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
if (insert_count % LSM_MERGE_CHECK_INTERVAL == 0) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE))
WT_ERR(EINTR);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
}
WT_ERR(src->get_key(src, &key));
dest->set_key(dest, &key);
WT_ERR(src->get_value(src, &value));
dest->set_value(dest, &value);
WT_ERR(dest->insert(dest));
if (create_bloom)
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, insert_count % LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Bloom size for %" PRIu64 " has %" PRIu64 " items inserted.",
record_count, insert_count));
/*
* Closing and syncing the files can take a while. Set the
* merge_syncing field so that compact knows it is still in
* progress.
*/
(void)__wt_atomic_add32(&lsm_tree->merge_syncing, 1);
in_sync = true;
/*
* We've successfully created the new chunk. Now install it. We need
* to ensure that the NO_CACHE flag is cleared and the bloom filter
* is closed (even if a step fails), so track errors but don't return
* until we've cleaned up.
*/
WT_TRET(src->close(src));
WT_TRET(dest->close(dest));
src = dest = NULL;
F_CLR(session, WT_SESSION_NO_CACHE);
/*
* We're doing advisory reads to fault the new trees into cache.
* Don't block if the cache is full: our next unit of work may be to
* discard some trees to free space.
*/
F_SET(session, WT_SESSION_NO_EVICTION);
if (create_bloom) {
if (ret == 0)
WT_TRET(__wt_bloom_finalize(bloom));
/*
* Read in a key to make sure the Bloom filters btree handle is
* open before it becomes visible to application threads.
* Otherwise application threads will stall while it is opened
* and internal pages are read into cache.
*/
if (ret == 0) {
WT_CLEAR(key);
WT_TRET_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
}
WT_TRET(__wt_bloom_close(bloom));
bloom = NULL;
}
WT_ERR(ret);
/*
* Open a handle on the new chunk before application threads attempt
* to access it, opening it pre-loads internal pages into the file
* system cache.
*/
cfg[1] = "checkpoint=" WT_CHECKPOINT;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
WT_TRET(dest->close(dest));
dest = NULL;
++lsm_tree->merge_progressing;
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
in_sync = false;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
/*
* Check whether we raced with another merge, and adjust the chunk
* array offset as necessary.
*/
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++)
if (lsm_tree->chunk[start_chunk]->id == start_id)
break;
/*
* It is safe to error out here - since the update can only fail
* prior to making updates to the tree.
*/
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, start_chunk, nchunks, chunk));
if (create_bloom)
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
chunk->count = insert_count;
chunk->generation = generation;
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
/*
* We have no current way of continuing if the metadata update fails,
* so we will panic in that case. Put some effort into cleaning up
* after ourselves here - so things have a chance of shutting down.
*
* Any errors that happened after the tree was locked are
* fatal - we can't guarantee the state of the tree.
*/
if ((ret = __wt_lsm_meta_write(session, lsm_tree)) != 0)
WT_PANIC_ERR(session, ret, "Failed finalizing LSM merge");
lsm_tree->dsk_gen++;
/* Update the throttling while holding the tree lock. */
__wt_lsm_tree_throttle(session, lsm_tree, true);
/* Schedule a pass to discard old chunks */
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_DROP, 0, lsm_tree));
err: if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (in_sync)
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
if (src != NULL)
WT_TRET(src->close(src));
if (dest != NULL)
WT_TRET(dest->close(dest));
if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
if (ret != 0 && created_chunk) {
/* Drop the newly-created files on error. */
if (chunk->uri != NULL) {
WT_WITH_SCHEMA_LOCK(session, tret =
__wt_schema_drop(session, chunk->uri, drop_cfg));
WT_TRET(tret);
}
if (create_bloom && chunk->bloom_uri != NULL) {
WT_WITH_SCHEMA_LOCK(session,
tret = __wt_schema_drop(
session, chunk->bloom_uri, drop_cfg));
WT_TRET(tret);
}
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, chunk);
if (ret == EINTR)
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge aborted due to close"));
else
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge failed with %s",
__wt_strerror(session, ret, NULL, 0)));
}
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __wt_lsm_free_chunks --
* Try to drop chunks from the tree that are no longer required.
*/
int
__wt_lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i, skipped;
int drop_ret;
bool flush_metadata;
flush_metadata = false;
if (lsm_tree->nold_chunks == 0)
return (0);
/*
* Make sure only a single thread is freeing the old chunk array
* at any time.
*/
if (!__wt_atomic_cas32(&lsm_tree->freeing_old_chunks, 0, 1))
return (0);
/*
* Take a copy of the current state of the LSM tree and look for chunks
* to drop. We do it this way to avoid holding the LSM tree lock while
* doing I/O or waiting on the schema lock.
*
* This is safe because only one thread will be in this function at a
* time. Merges may complete concurrently, and the old_chunks array
* may be extended, but we shuffle down the pointers each time we free
* one to keep the non-NULL slots at the beginning of the array.
*/
WT_CLEAR(cookie);
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, true));
for (i = skipped = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
WT_ASSERT(session, chunk != NULL);
/* Skip the chunk if another worker is using it. */
if (chunk->refcnt > 1) {
++skipped;
continue;
}
/*
* Don't remove files if a hot backup is in progress.
*
* The schema lock protects the set of live files, this check
* prevents us from removing a file that hot backup already
* knows about.
*/
if (S2C(session)->hot_backup)
break;
/*
* Drop any bloom filters and chunks we can. Don't try to drop
* a chunk if the bloom filter drop fails.
* An EBUSY return indicates that a cursor is still open in
* the tree - move to the next chunk in that case.
* An ENOENT return indicates that the LSM tree metadata was
* out of sync with the on disk state. Update the
* metadata to match in that case.
*/
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
drop_ret = __lsm_drop_file(session, chunk->bloom_uri);
if (drop_ret == EBUSY) {
++skipped;
continue;
} else if (drop_ret != ENOENT)
WT_ERR(drop_ret);
flush_metadata = true;
F_CLR(chunk, WT_LSM_CHUNK_BLOOM);
}
if (chunk->uri != NULL) {
drop_ret = __lsm_drop_file(session, chunk->uri);
if (drop_ret == EBUSY) {
++skipped;
continue;
} else if (drop_ret != ENOENT)
WT_ERR(drop_ret);
flush_metadata = true;
}
/* Lock the tree to clear out the old chunk information. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
/*
* The chunk we are looking at should be the first one in the
* tree that we haven't already skipped over.
*/
WT_ASSERT(session, lsm_tree->old_chunks[skipped] == chunk);
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, lsm_tree->old_chunks[skipped]);
/* Shuffle down to keep all occupied slots at the beginning. */
if (--lsm_tree->nold_chunks > skipped) {
memmove(lsm_tree->old_chunks + skipped,
lsm_tree->old_chunks + skipped + 1,
(lsm_tree->nold_chunks - skipped) *
sizeof(WT_LSM_CHUNK *));
lsm_tree->old_chunks[lsm_tree->nold_chunks] = NULL;
}
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
/*
* Clear the chunk in the cookie so we don't attempt to
* decrement the reference count.
*/
cookie.chunk_array[i] = NULL;
}
err: /* Flush the metadata unless the system is in panic */
if (flush_metadata && ret != WT_PANIC) {
WT_TRET(__wt_lsm_tree_writelock(session, lsm_tree));
WT_TRET(__wt_lsm_meta_write(session, lsm_tree));
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
}
__lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
lsm_tree->freeing_old_chunks = 0;
/* Returning non-zero means there is no work to do. */
if (!flush_metadata)
WT_TRET(WT_NOTFOUND);
return (ret);
}
/*
* __wt_lsm_tree_switch --
* Switch to a new in-memory tree.
*/
int
__wt_lsm_tree_switch(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
uint32_t nchunks, new_id;
int first_switch;
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
nchunks = lsm_tree->nchunks;
first_switch = nchunks == 0 ? 1 : 0;
/*
* Check if a switch is still needed: we may have raced while waiting
* for a lock.
*/
chunk = NULL;
if (!first_switch &&
(chunk = lsm_tree->chunk[nchunks - 1]) != NULL &&
!F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
goto err;
/* Set the switch transaction in the previous chunk, if necessary. */
if (chunk != NULL && chunk->switch_txn == WT_TXN_NONE)
chunk->switch_txn = __wt_txn_new_id(session);
/* Update the throttle time. */
__wt_lsm_tree_throttle(session, lsm_tree, 0);
new_id = WT_ATOMIC_ADD4(lsm_tree->last, 1);
WT_ERR(__wt_realloc_def(session, &lsm_tree->chunk_alloc,
nchunks + 1, &lsm_tree->chunk));
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Tree %s switch to: %" PRIu32 ", checkpoint throttle %ld, "
"merge throttle %ld", lsm_tree->name,
new_id, lsm_tree->ckpt_throttle, lsm_tree->merge_throttle));
WT_ERR(__wt_calloc_def(session, 1, &chunk));
chunk->id = new_id;
chunk->switch_txn = WT_TXN_NONE;
lsm_tree->chunk[lsm_tree->nchunks++] = chunk;
WT_ERR(__wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
F_CLR(lsm_tree, WT_LSM_TREE_NEED_SWITCH);
++lsm_tree->dsk_gen;
lsm_tree->modified = 1;
err: WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
/*
* Errors that happen during a tree switch leave the tree in a state
* where we can't make progress. Error out of WiredTiger.
*/
if (ret != 0)
WT_PANIC_RET(session, ret, "Failed doing LSM switch");
else if (!first_switch)
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_FLUSH, 0, lsm_tree));
return (ret);
}
/*
* __wt_lsm_tree_rename --
* Rename an LSM tree.
*/
int
__wt_lsm_tree_rename(WT_SESSION_IMPL *session,
const char *olduri, const char *newuri, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
const char *old;
u_int i;
int locked;
old = NULL;
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, olduri, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_ERR(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = 1;
/* Set the new name. */
WT_ERR(__lsm_tree_set_name(session, lsm_tree, newuri));
/* 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, &chunk->uri));
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, &chunk->bloom_uri));
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
WT_ERR(__wt_schema_rename(
session, old, chunk->uri, cfg));
__wt_free(session, old);
}
}
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
locked = 0;
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
WT_ERR(__wt_metadata_remove(session, olduri));
err: if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
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);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
const char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_search(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "collator", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->collator_name));
WT_ERR(__wt_config_gets(session, cfg, "lsm.auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_max", &cval));
lsm_tree->chunk_max = (uint64_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_size", &cval));
lsm_tree->chunk_size = (uint64_t)cval.val;
if (lsm_tree->chunk_size > lsm_tree->chunk_max)
WT_ERR_MSG(session, EINVAL,
"Chunk size (chunk_size) must be smaller than or equal to "
"the maximum chunk size (chunk_max)");
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
lsm_tree->merge_min = lsm_tree->merge_max / 2;
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_threads", &cval));
lsm_tree->merge_threads = (uint32_t)cval.val;
/* Sanity check that api_data.py is in sync with lsm.h */
WT_ASSERT(session, lsm_tree->merge_threads <= WT_LSM_MAX_WORKERS);
/*
* Set up the config for each chunk. If possible, avoid high latencies
* from fsync by flushing the cache every 8MB (will be overridden by
* any application setting).
*/
tmpconfig = "";
#ifdef HAVE_SYNC_FILE_RANGE
if (!S2C(session)->direct_io)
tmpconfig = "os_cache_dirty_max=8MB,";
#endif
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s%s,key_format=u,value_format=u", tmpconfig, config));
lsm_tree->file_config = __wt_buf_steal(session, buf, NULL);
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_search(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "collator", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->collator_name));
WT_ERR(__wt_config_gets(session, cfg, "lsm.auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_max", &cval));
lsm_tree->chunk_max = (uint64_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_size", &cval));
lsm_tree->chunk_size = (uint64_t)cval.val;
if (lsm_tree->chunk_size > lsm_tree->chunk_max)
WT_ERR_MSG(session, EINVAL,
"Chunk size (chunk_size) must be smaller than or equal to "
"the maximum chunk size (chunk_max)");
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_min", &cval));
lsm_tree->merge_min = (uint32_t)cval.val;
if (lsm_tree->merge_min > lsm_tree->merge_max)
WT_ERR_MSG(session, EINVAL,
"LSM merge_min must be less than or equal to merge_max");
/*
* Set up the config for each chunk.
*
* Make the memory_page_max double the chunk size, so application
* threads don't immediately try to force evict the chunk when the
* worker thread clears the NO_EVICTION flag.
*/
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u,memory_page_max=%" PRIu64,
config, 2 * lsm_tree->chunk_max));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &lsm_tree->file_config));
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __lsm_free_chunks --
* Try to drop chunks from the tree that are no longer required.
*/
static int
__lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i, skipped;
int progress;
/*
* Take a copy of the current state of the LSM tree and look for chunks
* to drop. We do it this way to avoid holding the LSM tree lock while
* doing I/O or waiting on the schema lock.
*
* This is safe because only one thread will be in this function at a
* time (the first merge thread). Merges may complete concurrently,
* and the old_chunks array may be extended, but we shuffle down the
* pointers each time we free one to keep the non-NULL slots at the
* beginning of the array.
*/
WT_CLEAR(cookie);
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, 1));
for (i = skipped = 0, progress = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
WT_ASSERT(session, chunk != NULL);
/* Skip the chunk if another worker is using it. */
if (chunk->refcnt > 1) {
++skipped;
continue;
}
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM)) {
/*
* An EBUSY return is acceptable - a cursor may still
* be positioned on this old chunk.
*/
if ((ret = __lsm_drop_file(
session, chunk->bloom_uri)) == EBUSY) {
WT_VERBOSE_ERR(session, lsm,
"LSM worker bloom drop busy: %s.",
chunk->bloom_uri);
++skipped;
continue;
} else
WT_ERR(ret);
F_CLR_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM);
}
if (chunk->uri != NULL) {
/*
* An EBUSY return is acceptable - a cursor may still
* be positioned on this old chunk.
*/
if ((ret = __lsm_drop_file(
session, chunk->uri)) == EBUSY) {
WT_VERBOSE_ERR(session, lsm,
"LSM worker drop busy: %s.",
chunk->uri);
++skipped;
continue;
} else
WT_ERR(ret);
}
progress = 1;
/* Lock the tree to clear out the old chunk information. */
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
/*
* The chunk we are looking at should be the first one in the
* tree that we haven't already skipped over.
*/
WT_ASSERT(session, lsm_tree->old_chunks[skipped] == chunk);
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, lsm_tree->old_chunks[skipped]);
/* Shuffle down to keep all occupied slots at the beginning. */
if (--lsm_tree->nold_chunks > skipped) {
memmove(lsm_tree->old_chunks + skipped,
lsm_tree->old_chunks + skipped + 1,
(lsm_tree->nold_chunks - skipped) *
sizeof(WT_LSM_CHUNK *));
lsm_tree->old_chunks[lsm_tree->nold_chunks] = NULL;
}
/*
* Clear the chunk in the cookie so we don't attempt to
* decrement the reference count.
*/
cookie.chunk_array[i] = NULL;
/*
* Update the metadata. We used to try to optimize by only
* updating the metadata once at the end, but the error
* handling is not straightforward.
*/
WT_TRET(__wt_lsm_meta_write(session, lsm_tree));
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
}
err: __lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
/* Returning non-zero means there is no work to do. */
if (!progress)
WT_TRET(WT_NOTFOUND);
return (ret);
}
/*
* __lsm_bloom_create --
* Create a bloom filter for a chunk of the LSM tree that has been
* checkpointed but not yet been merged.
*/
static int
__lsm_bloom_create(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk, u_int chunk_off)
{
WT_BLOOM *bloom;
WT_CURSOR *src;
WT_DECL_RET;
WT_ITEM buf, key;
WT_SESSION *wt_session;
uint64_t insert_count;
int exist;
/*
* Normally, the Bloom URI is populated when the chunk struct is
* allocated. After an open, however, it may not have been.
* Deal with that here.
*/
if (chunk->bloom_uri == NULL) {
WT_CLEAR(buf);
WT_RET(__wt_lsm_tree_bloom_name(
session, lsm_tree, chunk->id, &buf));
chunk->bloom_uri = __wt_buf_steal(session, &buf, NULL);
}
/*
* Drop the bloom filter first - there may be some content hanging over
* from an aborted merge or checkpoint.
*/
wt_session = &session->iface;
WT_RET(__wt_exist(session, chunk->bloom_uri + strlen("file:"), &exist));
if (exist)
WT_RET(wt_session->drop(wt_session, chunk->bloom_uri, "force"));
bloom = NULL;
/*
* This is merge-like activity, and we don't want compacts to give up
* because we are creating a bunch of bloom filters before merging.
*/
++lsm_tree->merge_progressing;
WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config, chunk->count,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
/* Open a special merge cursor just on this chunk. */
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, chunk_off, chunk->id, 1));
F_SET(session, WT_SESSION_NO_CACHE);
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
WT_ERR(src->get_key(src, &key));
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_TRET(src->close(src));
WT_TRET(__wt_bloom_finalize(bloom));
WT_ERR(ret);
F_CLR(session, WT_SESSION_NO_CACHE);
/* Load the new Bloom filter into cache. */
WT_CLEAR(key);
WT_ERR_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
WT_VERBOSE_ERR(session, lsm,
"LSM worker created bloom filter %s. "
"Expected %" PRIu64 " items, got %" PRIu64,
chunk->bloom_uri, chunk->count, insert_count);
/* Ensure the bloom filter is in the metadata. */
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
F_SET_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM bloom worker metadata write");
err: if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
F_CLR(session, WT_SESSION_NO_CACHE);
return (ret);
}
/*
* __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);
}
/*
* __wt_lsm_checkpoint_chunk --
* Flush a single LSM chunk to disk.
*/
int
__wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_DECL_RET;
WT_TXN_ISOLATION saved_isolation;
/*
* 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(chunk, WT_LSM_CHUNK_ONDISK) &&
!F_ISSET(chunk, WT_LSM_CHUNK_STABLE) &&
!chunk->evicted) {
if ((ret = __lsm_discard_handle(
session, chunk->uri, NULL)) == 0)
chunk->evicted = 1;
else if (ret == EBUSY)
ret = 0;
else
WT_RET_MSG(session, ret, "discard handle");
}
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s already on disk",
chunk->uri));
return (0);
}
/* Stop if a running transaction needs the chunk. */
__wt_txn_update_oldest(session);
if (chunk->switch_txn == WT_TXN_NONE ||
!__wt_txn_visible_all(session, chunk->switch_txn)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s: running transaction, return",
chunk->uri));
return (0);
}
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker flushing %s",
chunk->uri));
/*
* Flush the file before checkpointing: this is the expensive part in
* terms of I/O.
*
* 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.
*/
if ((ret = __wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0)) == 0) {
saved_isolation = session->txn.isolation;
session->txn.isolation = TXN_ISO_EVICTION;
ret = __wt_cache_op(session, NULL, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
WT_TRET(__wt_session_release_btree(session));
}
WT_RET(ret);
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s",
chunk->uri));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, chunk->uri,
__wt_checkpoint, NULL, NULL, 0));
if (ret != 0)
WT_RET_MSG(session, ret, "LSM checkpoint");
/* Now the file is written, get the chunk size. */
WT_RET(__wt_lsm_tree_set_chunk_size(session, chunk));
/* Update the flush timestamp to help track ongoing progress. */
WT_RET(__wt_epoch(session, &lsm_tree->last_flush_ts));
/* Lock the tree, mark the chunk as on disk and update the metadata. */
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
F_SET(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, 1);
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (ret != 0)
WT_RET_MSG(session, ret, "LSM metadata write");
/*
* Clear the no-eviction 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_RET(__wt_session_get_btree(session, chunk->uri, NULL, NULL, 0));
__wt_btree_evictable(session, 1);
WT_RET(__wt_session_release_btree(session));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointed %s",
chunk->uri));
/* Schedule a bloom filter create for our newly flushed chunk. */
if (!FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF))
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_BLOOM, 0, lsm_tree));
else
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
return (0);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
const char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_read(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_RET(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "lsm_auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_newest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST);
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST) ||
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST)))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_chunk_size", &cval));
lsm_tree->chunk_size = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_threads", &cval));
lsm_tree->merge_threads = (uint32_t)cval.val;
/* Sanity check that api_data.py is in sync with lsm.h */
WT_ASSERT(session, lsm_tree->merge_threads <= WT_LSM_MAX_WORKERS);
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u", config));
lsm_tree->file_config = __wt_buf_steal(session, buf, NULL);
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
