/*
* __wt_bloom_open --
* Open a Bloom filter object for use by a single session. The filter must
* have been created and finalized.
*/
int
__wt_bloom_open(WT_SESSION_IMPL *session,
const char *uri, uint32_t factor, uint32_t k,
WT_CURSOR *owner, WT_BLOOM **bloomp)
{
WT_BLOOM *bloom;
WT_CURSOR *c;
WT_DECL_RET;
uint64_t size;
WT_RET(__bloom_init(session, uri, NULL, &bloom));
WT_ERR(__bloom_open_cursor(bloom, owner));
c = bloom->c;
/* Find the largest key, to get the size of the filter. */
WT_ERR(c->prev(c));
WT_ERR(c->get_key(c, &size));
WT_ERR(c->reset(c));
WT_ERR(__bloom_setup(bloom, 0, size, factor, k));
*bloomp = bloom;
return (0);
err: (void)__wt_bloom_close(bloom);
return (ret);
}
/*
* __clsm_close_cursors --
* Close any btree cursors that are not needed.
*/
static int
__clsm_close_cursors(WT_CURSOR_LSM *clsm, u_int start, u_int end)
{
WT_BLOOM *bloom;
WT_CURSOR *c;
u_int i;
if (clsm->cursors == NULL || clsm->nchunks == 0)
return (0);
/*
* Walk the cursors, closing any we don't need. Note that the exit
* condition here is special, don't use WT_FORALL_CURSORS, and be
* careful with unsigned integer wrapping.
*/
for (i = start; i < end; i++) {
if ((c = (clsm)->cursors[i]) != NULL) {
clsm->cursors[i] = NULL;
WT_RET(c->close(c));
}
if ((bloom = clsm->blooms[i]) != NULL) {
clsm->blooms[i] = NULL;
WT_RET(__wt_bloom_close(bloom));
}
}
return (0);
}
/*
* __curjoin_close --
* WT_CURSOR::close for join cursors.
*/
static int
__curjoin_close(WT_CURSOR *cursor)
{
WT_CURSOR_JOIN *cjoin;
WT_CURSOR_JOIN_ENDPOINT *end;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
cjoin = (WT_CURSOR_JOIN *)cursor;
JOINABLE_CURSOR_API_CALL_PREPARE_ALLOWED(cursor, session, close, NULL);
err:
WT_TRET(__wt_schema_release_table(session, &cjoin->table));
/* This is owned by the table */
cursor->key_format = NULL;
if (cjoin->projection != NULL) {
__wt_free(session, cjoin->projection);
__wt_free(session, cursor->value_format);
}
for (entry = cjoin->entries, i = 0; i < cjoin->entries_next;
entry++, i++) {
if (entry->subjoin != NULL) {
F_CLR(&entry->subjoin->iface, WT_CURSTD_JOINED);
entry->subjoin->parent = NULL;
}
if (entry->main != NULL)
WT_TRET(entry->main->close(entry->main));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_OWN_BLOOM))
WT_TRET(__wt_bloom_close(entry->bloom));
for (end = &entry->ends[0];
end < &entry->ends[entry->ends_next]; end++) {
F_CLR(end->cursor, WT_CURSTD_JOINED);
if (F_ISSET(end, WT_CURJOIN_END_OWN_CURSOR))
WT_TRET(end->cursor->close(end->cursor));
}
__wt_free(session, entry->ends);
__wt_free(session, entry->repack_format);
}
if (cjoin->iter != NULL)
WT_TRET(__curjoin_iter_close_all(cjoin->iter));
if (cjoin->main != NULL)
WT_TRET(cjoin->main->close(cjoin->main));
__wt_free(session, cjoin->entries);
__wt_cursor_close(cursor);
API_END_RET(session, ret);
}
/*
* __wt_bloom_drop --
* Drop a Bloom filter, release any resources.
*/
int
__wt_bloom_drop(WT_BLOOM *bloom, const char *config)
{
WT_DECL_RET;
WT_SESSION *wt_session;
wt_session = (WT_SESSION *)bloom->session;
if (bloom->c != NULL) {
ret = bloom->c->close(bloom->c);
bloom->c = NULL;
}
WT_TRET(wt_session->drop(wt_session, bloom->uri, config));
WT_TRET(__wt_bloom_close(bloom));
return (ret);
}
/*
* __curjoin_close --
* WT_CURSOR::close for join cursors.
*/
static int
__curjoin_close(WT_CURSOR *cursor)
{
WT_CURSOR_JOIN *cjoin;
WT_CURSOR_JOIN_ENDPOINT *end;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
cjoin = (WT_CURSOR_JOIN *)cursor;
CURSOR_API_CALL(cursor, session, close, NULL);
__wt_schema_release_table(session, cjoin->table);
/* These are owned by the table */
cursor->internal_uri = NULL;
cursor->key_format = NULL;
if (cjoin->projection != NULL) {
__wt_free(session, cjoin->projection);
__wt_free(session, cursor->value_format);
}
for (entry = cjoin->entries, i = 0; i < cjoin->entries_next;
entry++, i++) {
if (entry->main != NULL)
WT_TRET(entry->main->close(entry->main));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_OWN_BLOOM))
WT_TRET(__wt_bloom_close(entry->bloom));
for (end = &entry->ends[0];
end < &entry->ends[entry->ends_next]; end++) {
F_CLR(end->cursor, WT_CURSTD_JOINED);
if (F_ISSET(end, WT_CURJOIN_END_OWN_KEY))
__wt_free(session, end->key.data);
}
__wt_free(session, entry->ends);
}
if (cjoin->iter != NULL)
WT_TRET(__curjoin_entry_iter_close(cjoin->iter));
__wt_free(session, cjoin->entries);
WT_TRET(__wt_cursor_close(cursor));
err: API_END_RET(session, ret);
}
/*
* __wt_bloom_create --
*
* Creates and configures a WT_BLOOM handle, allocates a bitstring in memory to
* use while populating the bloom filter.
*
* count - is the expected number of inserted items
* factor - is the number of bits to use per inserted item
* k - is the number of hash values to set or test per item
*/
int
__wt_bloom_create(
WT_SESSION_IMPL *session, const char *uri, const char *config,
uint64_t count, uint32_t factor, uint32_t k, WT_BLOOM **bloomp)
{
WT_BLOOM *bloom;
WT_DECL_RET;
WT_RET(__bloom_init(session, uri, config, &bloom));
WT_ERR(__bloom_setup(bloom, count, 0, factor, k));
WT_ERR(__bit_alloc(session, bloom->m, &bloom->bitstring));
*bloomp = bloom;
return (0);
err: (void)__wt_bloom_close(bloom);
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_ERR(__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_ERR(__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_TRET(__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);
}
void
run(void)
{
WT_BLOOM *bloomp;
WT_ITEM item;
WT_SESSION_IMPL *sess;
uint32_t fp, i;
int ret;
const char *uri = "file:my_bloom.bf";
/* Use the internal session handle to access private APIs. */
sess = (WT_SESSION_IMPL *)g.wt_session;
testutil_check(__wt_bloom_create(
sess, uri, NULL, g.c_ops, g.c_factor, g.c_k, &bloomp));
item.size = g.c_key_max;
for (i = 0; i < g.c_ops; i++) {
item.data = g.entries[i];
if ((ret = __wt_bloom_insert(bloomp, &item)) != 0)
testutil_die(ret, "__wt_bloom_insert: %" PRIu32, i);
}
testutil_check(__wt_bloom_finalize(bloomp));
for (i = 0; i < g.c_ops; i++) {
item.data = g.entries[i];
if ((ret = __wt_bloom_get(bloomp, &item)) != 0) {
fprintf(stderr,
"get failed at record: %" PRIu32 "\n", i);
testutil_die(ret, "__wt_bloom_get");
}
}
testutil_check(__wt_bloom_close(bloomp));
testutil_check(g.wt_session->checkpoint(g.wt_session, NULL));
testutil_check(__wt_bloom_open(
sess, uri, g.c_factor, g.c_k, NULL, &bloomp));
for (i = 0; i < g.c_ops; i++) {
item.data = g.entries[i];
testutil_check(__wt_bloom_get(bloomp, &item));
}
/*
* Try out some values we didn't insert - choose a different size to
* ensure the value doesn't overlap with existing values.
*/
item.size = g.c_key_max + 10;
item.data = dcalloc(item.size, 1);
memset((void *)item.data, 'a', item.size);
for (i = 0, fp = 0; i < g.c_ops; i++) {
((uint8_t *)item.data)[i % item.size] =
'a' + ((uint8_t)rand() % 26);
if ((ret = __wt_bloom_get(bloomp, &item)) == 0)
++fp;
if (ret != 0 && ret != WT_NOTFOUND)
testutil_die(ret, "__wt_bloom_get");
}
free((void *)item.data);
printf(
"Out of %" PRIu32 " ops, got %" PRIu32 " false positives, %.4f%%\n",
g.c_ops, fp, 100.0 * fp/g.c_ops);
testutil_check(__wt_bloom_drop(bloomp, NULL));
}
/*
* __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);
}
/*
* __curjoin_init_next --
* Initialize the cursor join when the next function is first called.
*/
static int
__curjoin_init_next(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
bool iterable)
{
WT_BLOOM *bloom;
WT_CURSOR *origcur;
WT_CURSOR_JOIN_ENDPOINT *end;
WT_CURSOR_JOIN_ENTRY *je, *jeend, *je2;
WT_DECL_RET;
size_t size;
uint32_t f, k;
char *mainbuf;
const char *def_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), NULL };
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
const char **config, *proj, *urimain;
mainbuf = NULL;
if (cjoin->entries_next == 0)
WT_RET_MSG(session, EINVAL,
"join cursor has not yet been joined with any other "
"cursors");
/* Get a consistent view of our subordinate cursors if appropriate. */
__wt_txn_cursor_op(session);
if (F_ISSET((WT_CURSOR *)cjoin, WT_CURSTD_RAW))
config = &raw_cfg[0];
else
config = &def_cfg[0];
urimain = cjoin->table->iface.name;
if ((proj = cjoin->projection) != NULL) {
size = strlen(urimain) + strlen(proj) + 1;
WT_ERR(__wt_calloc(session, size, 1, &mainbuf));
WT_ERR(__wt_snprintf(mainbuf, size, "%s%s", urimain, proj));
urimain = mainbuf;
}
WT_ERR(__wt_open_cursor(session, urimain, (WT_CURSOR *)cjoin, config,
&cjoin->main));
jeend = &cjoin->entries[cjoin->entries_next];
for (je = cjoin->entries; je < jeend; je++) {
if (je->subjoin != NULL) {
WT_ERR(__curjoin_init_next(session, je->subjoin,
iterable));
continue;
}
__wt_stat_join_init_single(&je->stats);
/*
* For a single compare=le/lt endpoint in any entry that may
* be iterated, construct a companion compare=ge endpoint
* that will actually be iterated.
*/
if (iterable && je->ends_next == 1 &&
F_ISSET(&je->ends[0], WT_CURJOIN_END_LT)) {
origcur = je->ends[0].cursor;
WT_ERR(__curjoin_insert_endpoint(session, je, 0, &end));
WT_ERR(__wt_open_cursor(session, origcur->uri,
(WT_CURSOR *)cjoin,
F_ISSET(origcur, WT_CURSTD_RAW) ? raw_cfg : def_cfg,
&end->cursor));
end->flags = WT_CURJOIN_END_GT | WT_CURJOIN_END_EQ |
WT_CURJOIN_END_OWN_CURSOR;
WT_ERR(end->cursor->next(end->cursor));
F_CLR(je, WT_CURJOIN_ENTRY_DISJUNCTION);
}
for (end = &je->ends[0]; end < &je->ends[je->ends_next];
end++)
WT_ERR(__curjoin_endpoint_init_key(session, je, end));
/*
* Do any needed Bloom filter initialization. Ignore Bloom
* filters for entries that will be iterated. They won't
* help since these entries either don't need an inclusion
* check or are doing any needed check during the iteration.
*/
if (!iterable && F_ISSET(je, WT_CURJOIN_ENTRY_BLOOM)) {
if (session->txn.isolation == WT_ISO_READ_UNCOMMITTED)
WT_ERR_MSG(session, EINVAL,
"join cursors with Bloom filters cannot be "
"used with read-uncommitted isolation");
if (je->bloom == NULL) {
/*
* Look for compatible filters to be shared,
* pick compatible numbers for bit counts
* and number of hashes.
*/
f = je->bloom_bit_count;
k = je->bloom_hash_count;
for (je2 = je + 1; je2 < jeend; je2++)
if (F_ISSET(je2,
WT_CURJOIN_ENTRY_BLOOM) &&
je2->count == je->count) {
f = WT_MAX(
je2->bloom_bit_count, f);
k = WT_MAX(
je2->bloom_hash_count, k);
}
je->bloom_bit_count = f;
je->bloom_hash_count = k;
WT_ERR(__wt_bloom_create(session, NULL,
NULL, je->count, f, k, &je->bloom));
F_SET(je, WT_CURJOIN_ENTRY_OWN_BLOOM);
WT_ERR(__curjoin_init_bloom(session, cjoin,
je, je->bloom));
/*
* Share the Bloom filter, making all
* config info consistent.
*/
for (je2 = je + 1; je2 < jeend; je2++)
if (F_ISSET(je2,
WT_CURJOIN_ENTRY_BLOOM) &&
je2->count == je->count) {
WT_ASSERT(session,
je2->bloom == NULL);
je2->bloom = je->bloom;
je2->bloom_bit_count = f;
je2->bloom_hash_count = k;
}
} else {
/*
* Create a temporary filter that we'll
* merge into the shared one. The Bloom
* parameters of the two filters must match.
*/
WT_ERR(__wt_bloom_create(session, NULL,
NULL, je->count, je->bloom_bit_count,
je->bloom_hash_count, &bloom));
WT_ERR(__curjoin_init_bloom(session, cjoin,
je, bloom));
WT_ERR(__wt_bloom_intersection(je->bloom,
bloom));
WT_ERR(__wt_bloom_close(bloom));
}
}
if (!F_ISSET(cjoin, WT_CURJOIN_DISJUNCTION))
iterable = false;
}
F_SET(cjoin, WT_CURJOIN_INITIALIZED);
err: __wt_free(session, mainbuf);
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);
}
/*
* __curjoin_init_iter --
* Initialize before any iteration.
*/
static int
__curjoin_init_iter(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin)
{
WT_BLOOM *bloom;
WT_DECL_RET;
WT_CURSOR_JOIN_ENTRY *je, *jeend, *je2;
WT_CURSOR_JOIN_ENDPOINT *end;
uint32_t f, k;
if (cjoin->entries_next == 0)
WT_RET_MSG(session, EINVAL,
"join cursor has not yet been joined with any other "
"cursors");
je = &cjoin->entries[0];
WT_RET(__curjoin_entry_iter_init(session, cjoin, je, &cjoin->iter));
jeend = &cjoin->entries[cjoin->entries_next];
for (je = cjoin->entries; je < jeend; je++) {
__wt_stat_join_init_single(&je->stats);
for (end = &je->ends[0]; end < &je->ends[je->ends_next];
end++)
WT_RET(__curjoin_endpoint_init_key(session, je, end));
/*
* The first entry is iterated as the 'outermost' cursor.
* For the common GE case, we don't have to test against
* the left reference key, we know it will be true since
* the btree is ordered.
*/
if (je == cjoin->entries && je->ends[0].flags ==
(WT_CURJOIN_END_GT | WT_CURJOIN_END_EQ))
F_SET(cjoin, WT_CURJOIN_SKIP_FIRST_LEFT);
if (F_ISSET(je, WT_CURJOIN_ENTRY_BLOOM)) {
if (je->bloom == NULL) {
/*
* Look for compatible filters to be shared,
* pick compatible numbers for bit counts
* and number of hashes.
*/
f = je->bloom_bit_count;
k = je->bloom_hash_count;
for (je2 = je + 1; je2 < jeend; je2++)
if (F_ISSET(je2,
WT_CURJOIN_ENTRY_BLOOM) &&
je2->count == je->count) {
f = WT_MAX(
je2->bloom_bit_count, f);
k = WT_MAX(
je2->bloom_hash_count, k);
}
je->bloom_bit_count = f;
je->bloom_hash_count = k;
WT_RET(__wt_bloom_create(session, NULL,
NULL, je->count, f, k, &je->bloom));
F_SET(je, WT_CURJOIN_ENTRY_OWN_BLOOM);
WT_RET(__curjoin_init_bloom(session, cjoin,
je, je->bloom));
/*
* Share the Bloom filter, making all
* config info consistent.
*/
for (je2 = je + 1; je2 < jeend; je2++)
if (F_ISSET(je2,
WT_CURJOIN_ENTRY_BLOOM) &&
je2->count == je->count) {
WT_ASSERT(session,
je2->bloom == NULL);
je2->bloom = je->bloom;
je2->bloom_bit_count = f;
je2->bloom_hash_count = k;
}
} else {
/*
* Create a temporary filter that we'll
* merge into the shared one. The Bloom
* parameters of the two filters must match.
*/
WT_RET(__wt_bloom_create(session, NULL,
NULL, je->count, je->bloom_bit_count,
je->bloom_hash_count, &bloom));
WT_RET(__curjoin_init_bloom(session, cjoin,
je, bloom));
WT_RET(__wt_bloom_intersection(je->bloom,
bloom));
WT_RET(__wt_bloom_close(bloom));
}
}
}
F_SET(cjoin, WT_CURJOIN_INITIALIZED);
return (ret);
}
