/*
* __wt_lsm_tree_setup_chunk --
* Initialize a chunk of an LSM tree.
*/
int
__wt_lsm_tree_setup_chunk(
WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_ITEM buf;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_drop), "force", NULL };
int exists;
WT_CLEAR(buf);
WT_RET(__wt_epoch(session, &chunk->create_ts));
WT_RET(__wt_lsm_tree_chunk_name(session, lsm_tree, chunk->id, &buf));
chunk->uri = __wt_buf_steal(session, &buf, NULL);
/*
* If the underlying file exists, drop the chunk first - there may be
* some content hanging over from an aborted merge or checkpoint.
*
* Don't do this for the very first chunk: we are called during
* WT_SESSION::create, and doing a drop inside there does interesting
* things with handle locks and metadata tracking. It can never have
* been the result of an interrupted merge, anyway.
*/
if (chunk->id > 1) {
WT_RET(__wt_exist(
session, chunk->uri + strlen("file:"), &exists));
if (exists)
WT_RET(__wt_schema_drop(session, chunk->uri, cfg));
}
return (__wt_schema_create(session, chunk->uri, lsm_tree->file_config));
}
/*
* __wt_lsm_tree_setup_chunk --
* Initialize a chunk of an LSM tree.
*/
int
__wt_lsm_tree_setup_chunk(
WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_ITEM buf;
const char *cfg[] = API_CONF_DEFAULTS(session, drop, "force");
WT_CLEAR(buf);
WT_RET(__wt_lsm_tree_chunk_name(session, lsm_tree, chunk->id, &buf));
chunk->uri = __wt_buf_steal(session, &buf, NULL);
/*
* Drop the chunk first - there may be some content hanging over from
* an aborted merge or checkpoint.
*
* Don't do this for the very first chunk: we are called during
* WT_SESSION::create, and doing a drop inside there does interesting
* things with handle locks and metadata tracking. It can never have
* been the result of an interrupted merge, anyway.
*/
if (chunk->id > 1)
WT_RET(__wt_schema_drop(session, chunk->uri, cfg));
return (__wt_schema_create(session, chunk->uri, lsm_tree->file_config));
}
/*
* __wt_lsm_tree_rename --
* Rename an LSM tree.
*/
int
__wt_lsm_tree_rename(WT_SESSION_IMPL *session,
const char *oldname, const char *newname, const char *cfg[])
{
WT_DECL_RET;
WT_ITEM buf;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
const char *old;
u_int i;
int locked;
old = NULL;
WT_CLEAR(buf);
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, oldname, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_ERR(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_ERR(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Set the new name. */
WT_ERR(__lsm_tree_set_name(session, lsm_tree, newname));
/* Rename the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
old = chunk->uri;
chunk->uri = NULL;
WT_ERR(__wt_lsm_tree_chunk_name(
session, lsm_tree, chunk->id, &buf));
chunk->uri = __wt_buf_steal(session, &buf, NULL);
WT_ERR(__wt_schema_rename(session, old, chunk->uri, cfg));
__wt_free(session, old);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
old = chunk->bloom_uri;
chunk->bloom_uri = NULL;
WT_ERR(__wt_lsm_tree_bloom_name(
session, lsm_tree, chunk->id, &buf));
chunk->bloom_uri = __wt_buf_steal(session, &buf, NULL);
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
WT_ERR(__wt_schema_rename(
session, old, chunk->uri, cfg));
__wt_free(session, old);
}
}
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
ret = __wt_lsm_meta_write(session, lsm_tree);
if (ret == 0)
ret = __wt_metadata_remove(session, oldname);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
if (old != NULL)
__wt_free(session, old);
/*
* Discard this LSM tree structure. The first operation on the renamed
* tree will create a new one.
*/
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __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[] = API_CONF_DEFAULTS(session, create, config);
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_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 == 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);
}
/*
* __wt_schema_open_colgroups --
* Open the column groups for a table.
*/
int
__wt_schema_open_colgroups(WT_SESSION_IMPL *session, WT_TABLE *table)
{
WT_COLGROUP *colgroup;
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_RET;
WT_DECL_ITEM(buf);
const char *cgconfig;
u_int i;
if (table->cg_complete)
return (0);
WT_RET(__wt_scr_alloc(session, 0, &buf));
colgroup = NULL;
WT_RET(__wt_config_subinit(session, &cparser, &table->cgconf));
/* Open each column group. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if (table->ncolgroups > 0)
WT_ERR(__wt_config_next(&cparser, &ckey, &cval));
else
WT_CLEAR(ckey);
/*
* Always open from scratch: we may have failed part of the way
* through opening a table, or column groups may have changed.
*/
if (table->cgroups[i] != NULL) {
__wt_schema_destroy_colgroup(
session, table->cgroups[i]);
table->cgroups[i] = NULL;
}
WT_ERR(__wt_schema_colgroup_name(session, table,
ckey.str, ckey.len, buf));
if ((ret = __wt_metadata_search(
session, buf->data, &cgconfig)) != 0) {
/* It is okay if the table is incomplete. */
if (ret == WT_NOTFOUND)
ret = 0;
goto err;
}
WT_ERR(__wt_calloc_def(session, 1, &colgroup));
colgroup->name = __wt_buf_steal(session, buf, NULL);
colgroup->config = cgconfig;
WT_ERR(__wt_config_getones(session,
colgroup->config, "columns", &colgroup->colconf));
WT_ERR(__wt_config_getones(
session, colgroup->config, "source", &cval));
WT_ERR(__wt_buf_fmt(session, buf, "%.*s",
(int)cval.len, cval.str));
colgroup->source = __wt_buf_steal(session, buf, NULL);
table->cgroups[i] = colgroup;
colgroup = NULL;
}
if (!table->is_simple) {
WT_ERR(__wt_table_check(session, table));
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, 1, buf));
table->plan = __wt_buf_steal(session, buf, NULL);
}
table->cg_complete = 1;
err: __wt_scr_free(&buf);
if (colgroup != NULL)
__wt_schema_destroy_colgroup(session, colgroup);
return (ret);
}
/*
* __wt_schema_open_table --
* Open a named table.
*/
int
__wt_schema_open_table(WT_SESSION_IMPL *session,
const char *name, size_t namelen, WT_TABLE **tablep)
{
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_ITEM buf;
WT_TABLE *table;
const char *tconfig;
char *tablename;
cursor = NULL;
table = NULL;
WT_CLEAR(buf);
WT_RET(__wt_buf_fmt(session, &buf, "table:%.*s", (int)namelen, name));
tablename = __wt_buf_steal(session, &buf, NULL);
WT_ERR(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, tablename);
WT_ERR(cursor->search(cursor));
WT_ERR(cursor->get_value(cursor, &tconfig));
WT_ERR(__wt_calloc_def(session, 1, &table));
table->name = tablename;
tablename = NULL;
WT_ERR(__wt_config_getones(session, tconfig, "columns", &cval));
WT_ERR(__wt_config_getones(session, tconfig, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->key_format));
WT_ERR(__wt_config_getones(session, tconfig, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len, &table->value_format));
WT_ERR(__wt_strdup(session, tconfig, &table->config));
/* Point to some items in the copy to save re-parsing. */
WT_ERR(__wt_config_getones(session, table->config,
"columns", &table->colconf));
/*
* Count the number of columns: tables are "simple" if the columns
* are not named.
*/
WT_ERR(__wt_config_subinit(session, &cparser, &table->colconf));
table->is_simple = 1;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
table->is_simple = 0;
if (ret != WT_NOTFOUND)
goto err;
/* Check that the columns match the key and value formats. */
if (!table->is_simple)
WT_ERR(__wt_schema_colcheck(session,
table->key_format, table->value_format, &table->colconf,
&table->nkey_columns, NULL));
WT_ERR(__wt_config_getones(session, table->config,
"colgroups", &table->cgconf));
/* Count the number of column groups. */
WT_ERR(__wt_config_subinit(session, &cparser, &table->cgconf));
table->ncolgroups = 0;
while ((ret = __wt_config_next(&cparser, &ckey, &cval)) == 0)
++table->ncolgroups;
if (ret != WT_NOTFOUND)
goto err;
WT_ERR(__wt_calloc_def(session, WT_COLGROUPS(table), &table->cgroups));
WT_ERR(__wt_schema_open_colgroups(session, table));
*tablep = table;
if (0) {
err: if (table != NULL)
__wt_schema_destroy_table(session, table);
}
if (cursor != NULL)
WT_TRET(cursor->close(cursor));
__wt_free(session, tablename);
return (ret);
}
/*
* ___open_index --
* Open an index.
*/
static int
__open_index(WT_SESSION_IMPL *session, WT_TABLE *table, WT_INDEX *idx)
{
WT_CONFIG colconf;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_ITEM(buf);
WT_DECL_ITEM(plan);
WT_DECL_RET;
u_int cursor_key_cols, i;
WT_ERR(__wt_scr_alloc(session, 0, &buf));
/* Get the data source from the index config. */
WT_ERR(__wt_config_getones(session, idx->config, "source", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->source = __wt_buf_steal(session, buf, NULL);
idx->need_value = WT_PREFIX_MATCH(idx->source, "lsm:");
WT_ERR(__wt_config_getones(session, idx->config, "key_format", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->key_format = __wt_buf_steal(session, buf, NULL);
/*
* The key format for an index is somewhat subtle: the application
* specifies a set of columns that it will use for the key, but the
* engine usually adds some hidden columns in order to derive the
* primary key. These hidden columns are part of the file's key.
*
* The file's key_format is stored persistently, we need to calculate
* the index cursor key format (which will usually omit some of those
* keys).
*/
WT_ERR(__wt_config_getones(
session, idx->config, "columns", &idx->colconf));
/* Start with the declared index columns. */
WT_ERR(__wt_config_subinit(session, &colconf, &idx->colconf));
cursor_key_cols = 0;
while ((ret = __wt_config_next(&colconf, &ckey, &cval)) == 0) {
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
++cursor_key_cols;
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/*
* Now add any primary key columns from the table that are not
* already part of the index key.
*/
WT_ERR(__wt_config_subinit(session, &colconf, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&colconf, &ckey, &cval)) == 0;
i++) {
/*
* If the primary key column is already in the secondary key,
* don't add it again.
*/
if (__wt_config_subgetraw(
session, &idx->colconf, &ckey, &cval) == 0)
continue;
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
WT_ERR(__wt_scr_alloc(session, 0, &plan));
WT_ERR(__wt_struct_plan(session, table, buf->data, buf->size, 0, plan));
idx->key_plan = __wt_buf_steal(session, plan, NULL);
/* Set up the cursor key format (the visible columns). */
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_struct_truncate(session,
idx->key_format, cursor_key_cols, buf));
idx->idxkey_format = __wt_buf_steal(session, buf, NULL);
/* By default, index cursor values are the table value columns. */
/* TODO Optimize to use index columns in preference to table lookups. */
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, 1, plan));
idx->value_plan = __wt_buf_steal(session, plan, NULL);
err: __wt_scr_free(&buf);
__wt_scr_free(&plan);
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);
}
/*
* __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);
}
