/*
* __truncate_dsrc --
* WT_SESSION::truncate for a data-source without a truncate operation.
*/
static int
__truncate_dsrc(WT_SESSION_IMPL *session, const char *uri)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *cfg[2];
/* Open a cursor and traverse the object, removing every entry. */
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = NULL;
WT_RET(__wt_open_cursor(session, uri, NULL, cfg, &cursor));
while ((ret = cursor->next(cursor)) == 0)
WT_ERR(cursor->remove(cursor));
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_DATA_INCR(session, cursor_truncate);
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __compact_checkpoint --
* Perform a checkpoint for compaction.
*/
static int
__compact_checkpoint(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
WT_TXN_GLOBAL *txn_global;
uint64_t txn_gen;
/*
* Force compaction checkpoints: we don't want to skip it because the
* work we need to have done is done in the underlying block manager.
*/
const char *checkpoint_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_checkpoint), "force=1", NULL };
/* Checkpoints take a lot of time, check if we've run out. */
WT_RET(__wt_session_compact_check_timeout(session));
if ((ret = __wt_txn_checkpoint(session, checkpoint_cfg, false)) == 0)
return (0);
WT_RET_BUSY_OK(ret);
/*
* If there's a checkpoint running, wait for it to complete, checking if
* we're out of time. If there's no checkpoint running or the checkpoint
* generation number changes, the checkpoint blocking us has completed.
*/
txn_global = &S2C(session)->txn_global;
for (txn_gen = __wt_gen(session, WT_GEN_CHECKPOINT);;) {
/*
* This loop only checks objects that are declared volatile,
* therefore no barriers are needed.
*/
if (!txn_global->checkpoint_running ||
txn_gen != __wt_gen(session, WT_GEN_CHECKPOINT))
break;
WT_RET(__wt_session_compact_check_timeout(session));
__wt_sleep(2, 0);
}
return (0);
}
/*
* __conn_btree_config_set --
* Set up a btree handle's configuration information.
*/
static int
__conn_btree_config_set(WT_SESSION_IMPL *session)
{
WT_DATA_HANDLE *dhandle;
WT_DECL_RET;
char *metaconf;
dhandle = session->dhandle;
/*
* Read the object's entry from the metadata file, we're done if we
* don't find one.
*/
if ((ret =
__wt_metadata_search(session, dhandle->name, &metaconf)) != 0) {
if (ret == WT_NOTFOUND)
ret = ENOENT;
WT_RET(ret);
}
/*
* The defaults are included because underlying objects have persistent
* configuration information stored in the metadata file. If defaults
* are included in the configuration, we can add new configuration
* strings without upgrading the metadata file or writing special code
* in case a configuration string isn't initialized, as long as the new
* configuration string has an appropriate default value.
*
* The error handling is a little odd, but be careful: we're holding a
* chunk of allocated memory in metaconf. If we fail before we copy a
* reference to it into the object's configuration array, we must free
* it, after the copy, we don't want to free it.
*/
WT_ERR(__wt_calloc_def(session, 3, &dhandle->cfg));
WT_ERR(__wt_strdup(
session, WT_CONFIG_BASE(session, file_meta), &dhandle->cfg[0]));
dhandle->cfg[1] = metaconf;
return (0);
err: __wt_free(session, metaconf);
return (ret);
}
/*
* __wt_las_create --
* Initialize the database's lookaside store.
*/
int
__wt_las_create(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
uint32_t session_flags;
const char *drop_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_drop), "force=true", NULL };
conn = S2C(session);
/* Read-only and in-memory configurations don't need the LAS table. */
if (F_ISSET(conn, WT_CONN_IN_MEMORY | WT_CONN_READONLY))
return (0);
/*
* Done at startup: we cannot do it on demand because we require the
* schema lock to create and drop the table, and it may not always be
* available.
*
* Discard any previous incarnation of the table.
*/
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_drop(session, WT_LAS_URI, drop_cfg));
WT_RET(ret);
/* Re-create the table. */
WT_RET(__wt_session_create(session, WT_LAS_URI, WT_LAS_FORMAT));
/*
* Open a shared internal session used to access the lookaside table.
* This session should never be tapped for eviction.
*/
session_flags = WT_SESSION_LOOKASIDE_CURSOR | WT_SESSION_NO_EVICTION;
WT_RET(__wt_open_internal_session(
conn, "lookaside table", true, session_flags, &conn->las_session));
/* Flag that the lookaside table has been created. */
F_SET(conn, WT_CONN_LAS_OPEN);
return (0);
}
/*
* __wt_metadata_cursor --
* Opens a cursor on the metadata.
*/
int
__wt_metadata_cursor(
WT_SESSION_IMPL *session, const char *config, WT_CURSOR **cursorp)
{
WT_DATA_HANDLE *saved_dhandle;
WT_DECL_RET;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_open_cursor), config, NULL };
saved_dhandle = session->dhandle;
WT_ERR(__wt_metadata_open(session));
WT_SET_BTREE_IN_SESSION(session, session->metafile);
WT_ERR(__wt_session_lock_btree(session, 0));
ret = __wt_curfile_create(session, NULL, cfg, 0, 0, cursorp);
/* Restore the caller's btree. */
err: session->dhandle = saved_dhandle;
return (ret);
}
/*
* __metadata_load_bulk --
* Create any bulk-loaded file stubs.
*/
static int
__metadata_load_bulk(WT_SESSION_IMPL *session)
{
WT_CURSOR *cursor;
WT_DECL_RET;
uint32_t allocsize;
bool exist;
const char *filecfg[] = {
WT_CONFIG_BASE(session, file_meta), NULL, NULL };
const char *key, *value;
/*
* If a file was being bulk-loaded during the hot backup, it will appear
* in the metadata file, but the file won't exist. Create on demand.
*/
WT_RET(__wt_metadata_cursor(session, &cursor));
while ((ret = cursor->next(cursor)) == 0) {
WT_ERR(cursor->get_key(cursor, &key));
if (!WT_PREFIX_SKIP(key, "file:"))
continue;
/* If the file exists, it's all good. */
WT_ERR(__wt_fs_exist(session, key, &exist));
if (exist)
continue;
/*
* If the file doesn't exist, assume it's a bulk-loaded file;
* retrieve the allocation size and re-create the file.
*/
WT_ERR(cursor->get_value(cursor, &value));
filecfg[1] = value;
WT_ERR(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
WT_ERR(__wt_block_manager_create(session, key, allocsize));
}
WT_ERR_NOTFOUND_OK(ret);
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
return (ret);
}
/*
* __curjoin_open_main --
* For the given index, open the main file with a projection
* that is the index keys.
*/
static int
__curjoin_open_main(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry)
{
WT_DECL_RET;
WT_INDEX *idx;
size_t len, newsize;
char *main_uri, *newformat;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
main_uri = newformat = NULL;
idx = entry->index;
newsize = strlen(cjoin->table->iface.name) + idx->colconf.len + 1;
WT_ERR(__wt_calloc(session, 1, newsize, &main_uri));
WT_ERR(__wt_snprintf(main_uri, newsize, "%s%.*s",
cjoin->table->iface.name, (int)idx->colconf.len, idx->colconf.str));
WT_ERR(__wt_open_cursor(session, main_uri,
(WT_CURSOR *)cjoin, raw_cfg, &entry->main));
if (idx->extractor == NULL) {
/*
* Add no-op padding so trailing 'u' formats are not
* transformed to 'U'. This matches what happens in
* the index. We don't do this when we have an
* extractor, extractors already use the padding
* byte trick.
*/
len = strlen(entry->main->value_format) + 3;
WT_ERR(__wt_calloc(session, len, 1, &newformat));
WT_ERR(__wt_snprintf(
newformat, len, "%s0x", entry->main->value_format));
__wt_free(session, entry->main->value_format);
entry->main->value_format = newformat;
newformat = NULL;
}
err: __wt_free(session, main_uri);
__wt_free(session, newformat);
return (ret);
}
/*
* __metadata_config --
* Return the default configuration information for the metadata file.
*/
static int
__metadata_config(WT_SESSION_IMPL *session, char **metaconfp)
{
WT_DECL_ITEM(buf);
WT_DECL_RET;
const char *cfg[] = { WT_CONFIG_BASE(session, file_meta), NULL, NULL };
*metaconfp = NULL;
/* Create a turtle file with default values. */
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"key_format=S,value_format=S,id=%d,version=(major=%d,minor=%d)",
WT_METAFILE_ID,
WT_BTREE_MAJOR_VERSION_MAX, WT_BTREE_MINOR_VERSION_MAX));
cfg[1] = buf->data;
ret = __wt_config_collapse(session, cfg, metaconfp);
err: __wt_scr_free(session, &buf);
return (ret);
}
/*
* __lsm_tree_open_check --
* Validate the configuration of an LSM tree.
*/
static int
__lsm_tree_open_check(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_CONFIG_ITEM cval;
uint64_t maxleafpage, required;
const char *cfg[] = { WT_CONFIG_BASE(
session, session_create), lsm_tree->file_config, NULL };
WT_RET(__wt_config_gets(session, cfg, "leaf_page_max", &cval));
maxleafpage = (uint64_t)cval.val;
/* Three chunks, plus one page for each participant in a merge. */
required = 3 * lsm_tree->chunk_size +
lsm_tree->merge_threads * (lsm_tree->merge_max * maxleafpage);
if (S2C(session)->cache_size < required)
WT_RET_MSG(session, EINVAL,
"The LSM configuration requires a cache size of at least %"
PRIu64 ". Configured size is %" PRIu64,
required, S2C(session)->cache_size);
return (0);
}
/*
* __wt_metadata_cursor_open --
* Opens a cursor on the metadata.
*/
int
__wt_metadata_cursor_open(
WT_SESSION_IMPL *session, const char *config, WT_CURSOR **cursorp)
{
WT_BTREE *btree;
WT_DECL_RET;
const char *open_cursor_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_open_cursor), config, NULL };
WT_WITHOUT_DHANDLE(session, ret = __wt_open_cursor(
session, WT_METAFILE_URI, NULL, open_cursor_cfg, cursorp));
WT_RET(ret);
/*
* Retrieve the btree from the cursor, rather than the session because
* we don't always switch the metadata handle in to the session before
* entering this function.
*/
btree = ((WT_CURSOR_BTREE *)(*cursorp))->btree;
/*
* Special settings for metadata: skew eviction so metadata almost
* always stays in cache and make sure metadata is logged if possible.
*
* Test before setting so updates can't race in subsequent opens (the
* first update is safe because it's single-threaded from
* wiredtiger_open).
*/
if (btree->evict_priority == 0)
WT_WITH_BTREE(session, btree,
__wt_evict_priority_set(session, WT_EVICT_INT_SKEW));
if (F_ISSET(btree, WT_BTREE_NO_LOGGING))
F_CLR(btree, WT_BTREE_NO_LOGGING);
/* The metadata file always uses checkpoint IDs in visibility checks. */
btree->include_checkpoint_txn = true;
return (0);
}
/*
* __wt_metadata_cursor_open --
* Opens a cursor on the metadata.
*/
int
__wt_metadata_cursor_open(
WT_SESSION_IMPL *session, const char *config, WT_CURSOR **cursorp)
{
WT_BTREE *btree;
WT_DECL_RET;
const char *open_cursor_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_open_cursor), config, NULL };
WT_WITHOUT_DHANDLE(session, ret = __wt_open_cursor(
session, WT_METAFILE_URI, NULL, open_cursor_cfg, cursorp));
WT_RET(ret);
/*
* Retrieve the btree from the cursor, rather than the session because
* we don't always switch the metadata handle in to the session before
* entering this function.
*/
btree = ((WT_CURSOR_BTREE *)(*cursorp))->btree;
/*
* Set special flags for the metadata file: eviction (the metadata file
* is in-memory and never evicted), logging (the metadata file is always
* logged if possible).
*
* Test flags before setting them so updates can't race in subsequent
* opens (the first update is safe because it's single-threaded from
* wiredtiger_open).
*/
if (!F_ISSET(btree, WT_BTREE_IN_MEMORY))
F_SET(btree, WT_BTREE_IN_MEMORY);
if (!F_ISSET(btree, WT_BTREE_NO_EVICTION))
F_SET(btree, WT_BTREE_NO_EVICTION);
if (F_ISSET(btree, WT_BTREE_NO_LOGGING))
F_CLR(btree, WT_BTREE_NO_LOGGING);
return (0);
}
/*
* __bloom_open_cursor --
* Open a cursor to read from a Bloom filter.
*/
static int
__bloom_open_cursor(WT_BLOOM *bloom, WT_CURSOR *owner)
{
WT_CURSOR *c;
WT_SESSION_IMPL *session;
const char *cfg[3];
if ((c = bloom->c) != NULL)
return (0);
session = bloom->session;
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = bloom->config;
cfg[2] = NULL;
c = NULL;
WT_RET(__wt_open_cursor(session, bloom->uri, owner, cfg, &c));
/* XXX Layering violation: bump the cache priority for Bloom filters. */
((WT_CURSOR_BTREE *)c)->btree->evict_priority = WT_EVICT_INT_SKEW;
bloom->c = c;
return (0);
}
/*
* __wt_las_cursor_open --
* Open a new lookaside table cursor.
*/
int
__wt_las_cursor_open(WT_SESSION_IMPL *session, WT_CURSOR **cursorp)
{
WT_BTREE *btree;
WT_DECL_RET;
const char *open_cursor_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_open_cursor), NULL };
WT_WITHOUT_DHANDLE(session, ret = __wt_open_cursor(
session, WT_LAS_URI, NULL, open_cursor_cfg, cursorp));
WT_RET(ret);
/*
* Retrieve the btree from the cursor, rather than the session because
* we don't always switch the LAS handle in to the session before
* entering this function.
*/
btree = ((WT_CURSOR_BTREE *)(*cursorp))->btree;
/*
* Set special flags for the lookaside table: the lookaside flag (used,
* for example, to avoid writing records during reconciliation), also
* turn off checkpoints and logging.
*
* Test flags before setting them so updates can't race in subsequent
* opens (the first update is safe because it's single-threaded from
* wiredtiger_open).
*/
if (!F_ISSET(btree, WT_BTREE_LOOKASIDE))
F_SET(btree, WT_BTREE_LOOKASIDE);
if (!F_ISSET(btree, WT_BTREE_NO_CHECKPOINT))
F_SET(btree, WT_BTREE_NO_CHECKPOINT);
if (!F_ISSET(btree, WT_BTREE_NO_LOGGING))
F_SET(btree, WT_BTREE_NO_LOGGING);
return (0);
}
/*
* __lsm_drop_file --
* Helper function to drop part of an LSM tree.
*/
static int
__lsm_drop_file(WT_SESSION_IMPL *session, const char *uri)
{
WT_DECL_RET;
const char *drop_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_drop), "remove_files=false", NULL };
/*
* We need to grab the schema lock to drop the file, so first try to
* make sure there is minimal work to freeing space in the cache. Only
* bother trying to discard the checkpoint handle: the in-memory handle
* should have been closed already.
*
* This will fail with EBUSY if the file is still in use.
*/
WT_WITH_HANDLE_LIST_LOCK(session,
ret = __lsm_discard_handle(session, uri, WT_CHECKPOINT));
WT_RET(ret);
/*
* Take the schema lock for the drop operation. Since __wt_schema_drop
* results in the hot backup lock being taken when it updates the
* metadata (which would be too late to prevent our drop).
*/
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_schema_drop(session, uri, drop_cfg));
if (ret == 0)
ret = __wt_remove(session, uri + strlen("file:"));
WT_RET(__wt_verbose(session, WT_VERB_LSM, "Dropped %s", uri));
if (ret == EBUSY || ret == ENOENT)
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker drop of %s failed with %d", uri, ret));
return (ret);
}
/*
* __wt_las_create --
* Initialize the database's lookaside store.
*/
int
__wt_las_create(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
const char *drop_cfg[] = {
WT_CONFIG_BASE(session, WT_SESSION_drop), "force=true", NULL };
conn = S2C(session);
/*
* Done at startup: we cannot do it on demand because we require the
* schema lock to create and drop the file, and it may not always be
* available.
*
* Open an internal session, used for the shared lookaside cursor.
*
* Sessions associated with a lookaside cursor should never be tapped
* for eviction.
*/
WT_RET(__wt_open_internal_session(
conn, "lookaside table", 1, 1, &conn->las_session));
session = conn->las_session;
F_SET(session, WT_SESSION_LOOKASIDE_CURSOR | WT_SESSION_NO_EVICTION);
/* Discard any previous incarnation of the file. */
WT_RET(__wt_session_drop(session, WT_LAS_URI, drop_cfg));
/* Re-create the file. */
WT_RET(__wt_session_create(session, WT_LAS_URI, WT_LAS_FORMAT));
/* Open the shared cursor. */
WT_WITHOUT_DHANDLE(session,
ret = __las_cursor_create(session, &conn->las_cursor));
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);
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, bool update, u_int start_chunk, uint32_t start_id)
{
WT_BTREE *btree;
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
u_int close_range_end, close_range_start;
bool locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
chunk = NULL;
locked = false;
lsm_tree = clsm->lsm_tree;
/*
* Ensure that any snapshot update has cursors on the right set of
* chunks to guarantee visibility is correct.
*/
if (update && txn->isolation == WT_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
/*
* Query operations need a full set of cursors. Overwrite cursors
* do queries in service of updates.
*/
if (!update || !F_ISSET(c, WT_CURSTD_OVERWRITE))
F_SET(clsm, WT_CLSM_OPEN_READ);
if (lsm_tree->nchunks == 0)
return (0);
ckpt_cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/*
* If the key is pointing to memory that is pinned by a chunk
* cursor, take a copy before closing cursors.
*/
if (F_ISSET(c, WT_CURSTD_KEY_INT))
WT_CURSOR_NEEDKEY(c);
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
WT_RET(__wt_lsm_tree_readlock(session, lsm_tree));
locked = true;
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id) {
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->switch_txn));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->switch_txn[ngood - 1] = chunk->switch_txn;
if (__wt_txn_visible_all(
session, chunk->switch_txn))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (clsm->cursors != NULL && ngood < clsm->nchunks) {
close_range_start = ngood;
close_range_end = clsm->nchunks;
} else if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0 ) {
close_range_start = 0;
close_range_end = WT_MIN(nchunks, clsm->nchunks);
if (close_range_end > nupdates)
close_range_end -= nupdates;
else
close_range_end = 0;
WT_ASSERT(session, ngood >= close_range_end);
} else {
close_range_end = 0;
close_range_start = 0;
}
if (close_range_end > close_range_start) {
saved_gen = lsm_tree->dsk_gen;
locked = false;
WT_ERR(__wt_lsm_tree_readunlock(session, lsm_tree));
WT_ERR(__clsm_close_cursors(
clsm, close_range_start, close_range_end));
WT_ERR(__wt_lsm_tree_readlock(session, lsm_tree));
locked = true;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->switch_txn[i] = chunk->switch_txn;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND && F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
/*
* Setup all cursors other than the primary to only do conflict
* checks on insert operations. This allows us to execute
* inserts on non-primary chunks as a way of checking for
* write conflicts with concurrent updates.
*/
if (i != nchunks - 1)
(*cp)->insert = __wt_curfile_update_check;
if (!F_ISSET(clsm, WT_CLSM_MERGE) &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL &&
!F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
chunk->switch_txn == WT_TXN_NONE) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
/*
* Disable eviction for the in-memory chunk. Also clear the
* bulk load flag here, otherwise eviction will be enabled by
* the first update.
*/
btree = ((WT_CURSOR_BTREE *)(primary))->btree;
if (btree->bulk_load_ok) {
btree->bulk_load_ok = false;
WT_WITH_BTREE(session, btree,
__wt_btree_evictable(session, false));
}
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err:
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_readunlock(session, lsm_tree));
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);
}
/*
* __create_file --
* Create a new 'file:' object.
*/
static int
__create_file(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_DECL_ITEM(val);
WT_DECL_RET;
uint32_t allocsize;
int is_metadata;
const char *filename, **p, *filecfg[] =
{ WT_CONFIG_BASE(session, file_meta), config, NULL, NULL };
char *fileconf;
fileconf = NULL;
is_metadata = strcmp(uri, WT_METAFILE_URI) == 0;
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
WT_RET_MSG(session, EINVAL, "Expected a 'file:' URI: %s", uri);
/* Check if the file already exists. */
if (!is_metadata && (ret =
__wt_metadata_search(session, uri, &fileconf)) != WT_NOTFOUND) {
if (exclusive)
WT_TRET(EEXIST);
goto err;
}
/* Sanity check the allocation size. */
WT_RET(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
/* Create the file. */
WT_ERR(__wt_block_manager_create(session, filename, allocsize));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, NULL, uri));
/*
* If creating an ordinary file, append the file ID and current version
* numbers to the passed-in configuration and insert the resulting
* configuration into the metadata.
*/
if (!is_metadata) {
WT_ERR(__wt_scr_alloc(session, 0, &val));
WT_ERR(__wt_buf_fmt(session, val,
"id=%" PRIu32 ",version=(major=%d,minor=%d)",
++S2C(session)->next_file_id,
WT_BTREE_MAJOR_VERSION_MAX, WT_BTREE_MINOR_VERSION_MAX));
for (p = filecfg; *p != NULL; ++p)
;
*p = val->data;
WT_ERR(__wt_config_collapse(session, filecfg, &fileconf));
WT_ERR(__wt_metadata_insert(session, uri, fileconf));
}
/*
* Open the file to check that it was setup correctly. We don't need to
* pass the configuration, we just wrote the collapsed configuration
* into the metadata file, and it's going to be read/used by underlying
* functions.
*
* Keep the handle exclusive until it is released at the end of the
* call, otherwise we could race with a drop.
*/
WT_ERR(__wt_session_get_btree(
session, uri, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_handle_lock(session, 1));
else
WT_ERR(__wt_session_release_btree(session));
err: __wt_scr_free(session, &val);
__wt_free(session, fileconf);
return (ret);
}
/*
* __create_table --
* Create a table.
*/
static int
__create_table(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG conf;
WT_CONFIG_ITEM cgkey, cgval, cval;
WT_DECL_RET;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, table_meta), config, NULL, NULL };
const char *tablename;
char *tableconf, *cgname;
size_t cgsize;
int ncolgroups;
cgname = NULL;
table = NULL;
tableconf = NULL;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "table:"))
return (EINVAL);
if ((ret = __wt_schema_get_table(session,
tablename, strlen(tablename), 0, &table)) == 0) {
__wt_schema_release_table(session, table);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_ERR(__wt_config_gets(session, cfg, "colgroups", &cval));
WT_ERR(__wt_config_subinit(session, &conf, &cval));
for (ncolgroups = 0;
(ret = __wt_config_next(&conf, &cgkey, &cgval)) == 0;
ncolgroups++)
;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_collapse(session, cfg, &tableconf));
if ((ret = __wt_metadata_insert(session, name, tableconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Attempt to open the table now to catch any errors. */
WT_ERR(__wt_schema_get_table(
session, tablename, strlen(tablename), 1, &table));
if (ncolgroups == 0) {
cgsize = strlen("colgroup:") + strlen(tablename) + 1;
WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
snprintf(cgname, cgsize, "colgroup:%s", tablename);
WT_ERR(__create_colgroup(session, cgname, exclusive, config));
}
if (0) {
err: if (table != NULL) {
WT_TRET(__wt_schema_remove_table(session, table));
table = NULL;
}
}
if (table != NULL)
__wt_schema_release_table(session, table);
__wt_free(session, cgname);
__wt_free(session, tableconf);
return (ret);
}
/*
* __create_index --
* Create an index.
*/
static int
__create_index(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG kcols, pkcols;
WT_CONFIG_ITEM ckey, cval, icols, kval;
WT_DECL_PACK_VALUE(pv);
WT_DECL_RET;
WT_ITEM confbuf, extra_cols, fmt, namebuf;
WT_PACK pack;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, index_meta), NULL, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *source, *sourceconf, *idxname, *tablename;
char *idxconf;
size_t tlen;
int have_extractor;
u_int i, npublic_cols;
sourceconf = NULL;
idxconf = NULL;
WT_CLEAR(confbuf);
WT_CLEAR(fmt);
WT_CLEAR(extra_cols);
WT_CLEAR(namebuf);
have_extractor = 0;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "index:"))
return (EINVAL);
idxname = strchr(tablename, ':');
if (idxname == NULL)
WT_RET_MSG(session, EINVAL, "Invalid index name, "
"should be <table name>:<index name>: %s", name);
tlen = (size_t)(idxname++ - tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, ret,
"Can't create an index for a non-existent table: %.*s",
(int)tlen, tablename);
if (table->is_simple)
WT_RET_MSG(session, EINVAL,
"%s requires a table with named columns", name);
if (__wt_config_getones(session, config, "source", &cval) == 0) {
WT_ERR(__wt_buf_fmt(session, &namebuf,
"%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_index_source(
session, table, idxname, config, &namebuf));
source = namebuf.data;
/* Add the source name to the index config before collapsing. */
WT_ERR(__wt_buf_catfmt(session, &confbuf,
",source=\"%s\"", source));
}
if (__wt_config_getones_none(
session, config, "extractor", &cval) == 0 && cval.len != 0) {
have_extractor = 1;
/* Custom extractors must supply a key format. */
if ((ret = __wt_config_getones(
session, config, "key_format", &kval)) != 0)
WT_ERR_MSG(session, EINVAL,
"%s: custom extractors require a key_format", name);
}
/* Calculate the key/value formats. */
WT_CLEAR(icols);
if (__wt_config_getones(session, config, "columns", &icols) != 0 &&
!have_extractor)
WT_ERR_MSG(session, EINVAL,
"%s: requires 'columns' configuration", name);
/*
* Count the public columns using the declared columns for normal
* indices or the key format for custom extractors.
*/
npublic_cols = 0;
if (!have_extractor) {
WT_ERR(__wt_config_subinit(session, &kcols, &icols));
while ((ret = __wt_config_next(&kcols, &ckey, &cval)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
} else {
WT_ERR(__pack_initn(session, &pack, kval.str, kval.len));
while ((ret = __pack_next(&pack, &pv)) == 0)
++npublic_cols;
WT_ERR_NOTFOUND_OK(ret);
}
/*
* 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 source's
* key_format, which we are calculating now, but not part of an index
* cursor's key_format.
*/
WT_ERR(__wt_config_subinit(session, &pkcols, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&pkcols, &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, &icols, &ckey, &cval) == 0) {
if (have_extractor)
WT_ERR_MSG(session, EINVAL,
"an index with a custom extractor may not "
"include primary key columns");
continue;
}
WT_ERR(__wt_buf_catfmt(
session, &extra_cols, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/* Index values are empty: all columns are packed into the index key. */
WT_ERR(__wt_buf_fmt(session, &fmt, "value_format=,key_format="));
if (have_extractor) {
WT_ERR(__wt_buf_catfmt(session, &fmt, "%.*s",
(int)kval.len, kval.str));
WT_CLEAR(icols);
}
/*
* Construct the index key format, or append the primary key columns
* for custom extractors.
*/
WT_ERR(__wt_struct_reformat(session, table,
icols.str, icols.len, (const char *)extra_cols.data, 0, &fmt));
/* Check for a record number index key, which makes no sense. */
WT_ERR(__wt_config_getones(session, fmt.data, "key_format", &cval));
if (cval.len == 1 && cval.str[0] == 'r')
WT_ERR_MSG(session, EINVAL,
"column-store index may not use the record number as its "
"index key");
WT_ERR(__wt_buf_catfmt(
session, &fmt, ",index_key_columns=%u", npublic_cols));
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
cfg[1] = sourceconf;
cfg[2] = confbuf.data;
WT_ERR(__wt_config_collapse(session, cfg, &idxconf));
if ((ret = __wt_metadata_insert(session, name, idxconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
/* Make sure that the configuration is valid. */
WT_ERR(__wt_schema_open_index(
session, table, idxname, strlen(idxname), NULL));
err: __wt_free(session, idxconf);
__wt_free(session, sourceconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &extra_cols);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
return (ret);
}
/*
* __create_colgroup --
* Create a column group.
*/
static int
__create_colgroup(WT_SESSION_IMPL *session,
const char *name, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_ITEM confbuf, fmt, namebuf;
WT_TABLE *table;
size_t tlen;
const char **cfgp, *cfg[4] =
{ WT_CONFIG_BASE(session, colgroup_meta), config, NULL, NULL };
const char *sourcecfg[] = { config, NULL, NULL };
const char *cgname, *source, *sourceconf, *tablename;
char *cgconf, *oldconf;
sourceconf = NULL;
cgconf = oldconf = NULL;
WT_CLEAR(fmt);
WT_CLEAR(confbuf);
WT_CLEAR(namebuf);
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "colgroup:"))
return (EINVAL);
cgname = strchr(tablename, ':');
if (cgname != NULL) {
tlen = (size_t)(cgname - tablename);
++cgname;
} else
tlen = strlen(tablename);
if ((ret =
__wt_schema_get_table(session, tablename, tlen, 1, &table)) != 0)
WT_RET_MSG(session, (ret == WT_NOTFOUND) ? ENOENT : ret,
"Can't create '%s' for non-existent table '%.*s'",
name, (int)tlen, tablename);
/* Make sure the column group is referenced from the table. */
if (cgname != NULL && (ret =
__wt_config_subgets(session, &table->cgconf, cgname, &cval)) != 0)
WT_ERR_MSG(session, EINVAL,
"Column group '%s' not found in table '%.*s'",
cgname, (int)tlen, tablename);
/* Find the first NULL entry in the cfg stack. */
for (cfgp = &cfg[1]; *cfgp; cfgp++)
;
/* Add the source to the colgroup config before collapsing. */
if (__wt_config_getones(
session, config, "source", &cval) == 0 && cval.len != 0) {
WT_ERR(__wt_buf_fmt(
session, &namebuf, "%.*s", (int)cval.len, cval.str));
source = namebuf.data;
} else {
WT_ERR(__wt_schema_colgroup_source(
session, table, cgname, config, &namebuf));
source = namebuf.data;
WT_ERR(__wt_buf_fmt(
session, &confbuf, "source=\"%s\"", source));
*cfgp++ = confbuf.data;
}
/* Calculate the key/value formats: these go into the source config. */
WT_ERR(__wt_buf_fmt(session, &fmt, "key_format=%s", table->key_format));
if (cgname == NULL)
WT_ERR(__wt_buf_catfmt
(session, &fmt, ",value_format=%s", table->value_format));
else {
if (__wt_config_getones(session, config, "columns", &cval) != 0)
WT_ERR_MSG(session, EINVAL,
"No 'columns' configuration for '%s'", name);
WT_ERR(__wt_buf_catfmt(session, &fmt, ",value_format="));
WT_ERR(__wt_struct_reformat(session,
table, cval.str, cval.len, NULL, 1, &fmt));
}
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_merge(session, sourcecfg, NULL, &sourceconf));
WT_ERR(__wt_schema_create(session, source, sourceconf));
WT_ERR(__wt_config_collapse(session, cfg, &cgconf));
if ((ret = __wt_metadata_insert(session, name, cgconf)) != 0) {
/*
* If the entry already exists in the metadata, we're done.
* This is an error for exclusive creates but okay otherwise.
*/
if (ret == WT_DUPLICATE_KEY)
ret = exclusive ? EEXIST : 0;
goto err;
}
WT_ERR(__wt_schema_open_colgroups(session, table));
err: __wt_free(session, cgconf);
__wt_free(session, sourceconf);
__wt_free(session, oldconf);
__wt_buf_free(session, &confbuf);
__wt_buf_free(session, &fmt);
__wt_buf_free(session, &namebuf);
__wt_schema_release_table(session, table);
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);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_ITEM curkey, curvalue;
size_t size;
u_int skip;
int cmp;
const char *uri;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
c = NULL;
skip = 0;
if (entry->index != NULL)
/*
* Open the raw index. We're avoiding any references
* to the main table, they may be expensive.
*/
uri = entry->index->source;
else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->iface.name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->iface.name));
uri = uribuf->data;
}
WT_ERR(__wt_open_cursor(session, uri, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax) {
if (F_ISSET(end, WT_CURJOIN_END_GT) ||
WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_EQ) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_GE)
skip = 1;
} else if (F_ISSET(end, WT_CURJOIN_END_LT)) {
if ((ret = c->next(c)) == WT_NOTFOUND)
goto done;
WT_ERR(ret);
} else
WT_PANIC_ERR(session, EINVAL,
"fatal error in join cursor position state");
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
entry->stats.iterated++;
if (entry->index != NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
WT_ERR(__wt_struct_repack(session,
c->key_format,
(entry->repack_format != NULL ?
entry->repack_format : entry->index->idxkey_format),
&c->key, &curkey));
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION)) {
/* if condition satisfied, insert immediately */
switch (WT_CURJOIN_END_RANGE(end)) {
case WT_CURJOIN_END_EQ:
if (cmp == 0)
goto insert;
break;
case WT_CURJOIN_END_GT:
if (cmp > 0) {
/* skip this check next time */
skip = entry->ends_next;
goto insert;
}
break;
case WT_CURJOIN_END_GE:
if (cmp >= 0)
goto insert;
break;
case WT_CURJOIN_END_LT:
if (cmp < 0)
goto insert;
break;
case WT_CURJOIN_END_LE:
if (cmp <= 0)
goto insert;
break;
}
} else if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
/*
* Either it's a disjunction that hasn't satisfied any
* condition, or it's a conjunction that has satisfied all
* conditions.
*/
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
goto advance;
insert:
if (entry->index != NULL) {
curvalue.data =
(unsigned char *)curkey.data + curkey.size;
WT_ASSERT(session, c->key.size > curkey.size);
curvalue.size = c->key.size - curkey.size;
}
else
WT_ERR(c->get_key(c, &curvalue));
__wt_bloom_insert(bloom, &curvalue);
entry->stats.bloom_insert++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
return (ret);
}
/*
* __curjoin_iter_set_entry --
* Set the current entry for an iterator.
*/
static int
__curjoin_iter_set_entry(WT_CURSOR_JOIN_ITER *iter, u_int entry_pos)
{
WT_CURSOR *c, *to_dup;
WT_CURSOR_JOIN *cjoin, *topjoin;
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_SESSION_IMPL *session;
size_t size;
const char *raw_cfg[] = { WT_CONFIG_BASE(
iter->session, WT_SESSION_open_cursor), "raw", NULL };
const char *def_cfg[] = { WT_CONFIG_BASE(
iter->session, WT_SESSION_open_cursor), NULL };
const char **config;
char *uri;
session = iter->session;
cjoin = iter->cjoin;
uri = NULL;
entry = iter->entry = &cjoin->entries[entry_pos];
iter->positioned = false;
iter->entry_pos = entry_pos;
iter->end_pos = 0;
iter->is_equal = (entry->ends_next == 1 &&
WT_CURJOIN_END_RANGE(&entry->ends[0]) == WT_CURJOIN_END_EQ);
iter->end_skip = (entry->ends_next > 0 &&
WT_CURJOIN_END_RANGE(&entry->ends[0]) == WT_CURJOIN_END_GE) ? 1 : 0;
iter->end_count = WT_MIN(1, entry->ends_next);
if (F_ISSET(cjoin, WT_CURJOIN_DISJUNCTION)) {
iter->entry_count = cjoin->entries_next;
if (iter->is_equal)
iter->end_count = entry->ends_next;
} else
iter->entry_count = 1;
WT_ASSERT(iter->session, iter->entry_pos < iter->entry_count);
entry->stats.iterated = 0;
if (entry->subjoin == NULL) {
for (topjoin = iter->cjoin; topjoin->parent != NULL;
topjoin = topjoin->parent)
;
to_dup = entry->ends[0].cursor;
if (F_ISSET((WT_CURSOR *)topjoin, WT_CURSTD_RAW))
config = &raw_cfg[0];
else
config = &def_cfg[0];
size = strlen(to_dup->internal_uri) + 3;
WT_ERR(__wt_calloc(session, size, 1, &uri));
WT_ERR(__wt_snprintf(uri, size, "%s()", to_dup->internal_uri));
if ((c = iter->cursor) == NULL || strcmp(c->uri, uri) != 0) {
iter->cursor = NULL;
if (c != NULL)
WT_ERR(c->close(c));
WT_ERR(__wt_open_cursor(session, uri,
(WT_CURSOR *)topjoin, config, &iter->cursor));
}
WT_ERR(__wt_cursor_dup_position(to_dup, iter->cursor));
} else if (iter->cursor != NULL) {
WT_ERR(iter->cursor->close(iter->cursor));
iter->cursor = NULL;
}
err: __wt_free(session, uri);
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);
}
/*
* wiredtiger_open --
* Main library entry point: open a new connection to a WiredTiger
* database.
*/
int
wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
const char *config, WT_CONNECTION **wt_connp)
{
static const WT_CONNECTION stdc = {
__conn_close,
__conn_reconfigure,
__conn_get_home,
__conn_configure_method,
__conn_is_new,
__conn_open_session,
__conn_load_extension,
__conn_add_data_source,
__conn_add_collator,
__conn_add_compressor,
__conn_add_extractor,
__conn_get_extension_api
};
static const struct {
const char *name;
uint32_t flag;
} *ft, directio_types[] = {
{ "data", WT_DIRECTIO_DATA },
{ "log", WT_DIRECTIO_LOG },
{ NULL, 0 }
};
WT_CONFIG subconfig;
WT_CONFIG_ITEM cval, skey, sval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_ITEM *cbuf, expath, exconfig;
WT_SESSION_IMPL *session;
const char *cfg[5];
int exist;
*wt_connp = NULL;
session = NULL;
cbuf = NULL;
WT_CLEAR(expath);
WT_CLEAR(exconfig);
WT_RET(__wt_library_init());
WT_RET(__wt_calloc_def(NULL, 1, &conn));
conn->iface = stdc;
/*
* Immediately link the structure into the connection structure list:
* the only thing ever looked at on that list is the database name,
* and a NULL value is fine.
*/
__wt_spin_lock(NULL, &__wt_process.spinlock);
TAILQ_INSERT_TAIL(&__wt_process.connqh, conn, q);
__wt_spin_unlock(NULL, &__wt_process.spinlock);
session = conn->default_session = &conn->dummy_session;
session->iface.connection = &conn->iface;
session->name = "wiredtiger_open";
__wt_event_handler_set(session, event_handler);
/* Remaining basic initialization of the connection structure. */
WT_ERR(__wt_connection_init(conn));
/* Check/set the configuration strings. */
WT_ERR(__wt_config_check(session,
WT_CONFIG_REF(session, wiredtiger_open), config, 0));
cfg[0] = WT_CONFIG_BASE(session, wiredtiger_open);
cfg[1] = config;
/* Leave space for optional additional configuration. */
cfg[2] = cfg[3] = cfg[4] = NULL;
/* Get the database home. */
WT_ERR(__conn_home(session, home, cfg));
/* Make sure no other thread of control already owns this database. */
WT_ERR(__conn_single(session, cfg));
/* Read the database-home configuration file. */
WT_ERR(__conn_config_file(session, cfg, &cbuf));
/* Read the environment variable configuration. */
WT_ERR(__conn_config_env(session, cfg));
WT_ERR(__wt_config_gets(session, cfg, "hazard_max", &cval));
conn->hazard_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "session_max", &cval));
conn->session_size = (uint32_t)cval.val + WT_NUM_INTERNAL_SESSIONS;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge", &cval));
if (cval.val)
F_SET(conn, WT_CONN_LSM_MERGE);
WT_ERR(__wt_config_gets(session, cfg, "sync", &cval));
if (cval.val)
F_SET(conn, WT_CONN_SYNC);
WT_ERR(__wt_config_gets(session, cfg, "transactional", &cval));
if (cval.val)
F_SET(conn, WT_CONN_TRANSACTIONAL);
/* Configure verbose flags. */
WT_ERR(__conn_verbose_config(session, cfg));
WT_ERR(__wt_conn_cache_pool_config(session, cfg));
WT_ERR(__wt_config_gets(session, cfg, "logging", &cval));
if (cval.val != 0)
WT_ERR(__wt_open(
session, WT_LOG_FILENAME, 1, 0, 0, &conn->log_fh));
/* Configure direct I/O and buffer alignment. */
WT_ERR(__wt_config_gets(session, cfg, "buffer_alignment", &cval));
if (cval.val == -1)
conn->buffer_alignment = WT_BUFFER_ALIGNMENT_DEFAULT;
else
conn->buffer_alignment = (size_t)cval.val;
#ifndef HAVE_POSIX_MEMALIGN
if (conn->buffer_alignment != 0)
WT_ERR_MSG(session, EINVAL,
"buffer_alignment requires posix_memalign");
#endif
/*
* Configuration: direct_io, mmap, statistics.
*/
WT_ERR(__wt_config_gets(session, cfg, "direct_io", &cval));
for (ft = directio_types; ft->name != NULL; ft++) {
ret = __wt_config_subgets(session, &cval, ft->name, &sval);
if (ret == 0) {
if (sval.val)
FLD_SET(conn->direct_io, ft->flag);
} else if (ret != WT_NOTFOUND)
goto err;
}
WT_ERR(__wt_config_gets(session, cfg, "mmap", &cval));
conn->mmap = cval.val == 0 ? 0 : 1;
WT_ERR(__wt_config_gets(session, cfg, "statistics", &cval));
conn->statistics = cval.val == 0 ? 0 : 1;
/* Load any extensions referenced in the config. */
WT_ERR(__wt_config_gets(session, cfg, "extensions", &cval));
WT_ERR(__wt_config_subinit(session, &subconfig, &cval));
while ((ret = __wt_config_next(&subconfig, &skey, &sval)) == 0) {
WT_ERR(__wt_buf_fmt(
session, &expath, "%.*s", (int)skey.len, skey.str));
if (sval.len > 0)
WT_ERR(__wt_buf_fmt(session, &exconfig,
"entry=%.*s\n", (int)sval.len, sval.str));
WT_ERR(conn->iface.load_extension(&conn->iface,
expath.data, (sval.len > 0) ? exconfig.data : NULL));
}
WT_ERR_NOTFOUND_OK(ret);
/*
* Open the connection; if that fails, the connection handle has been
* destroyed by the time the open function returns.
*/
if ((ret = __wt_connection_open(conn, cfg)) != 0) {
conn = NULL;
WT_ERR(ret);
}
/* Open the default session. */
WT_ERR(__wt_open_session(conn, 1, NULL, NULL, &conn->default_session));
session = conn->default_session;
/*
* Check on the turtle and metadata files, creating them if necessary
* (which avoids application threads racing to create the metadata file
* later).
*/
WT_ERR(__wt_meta_turtle_init(session, &exist));
if (!exist) {
/*
* We're single-threaded, but acquire the schema lock
* regardless: the lower level code checks that it is
* appropriately synchronized.
*/
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_create(session, WT_METADATA_URI, NULL));
WT_ERR(ret);
}
WT_ERR(__wt_metadata_open(session));
/* If there's a hot-backup file, load it. */
WT_ERR(__wt_metadata_load_backup(session));
STATIC_ASSERT(offsetof(WT_CONNECTION_IMPL, iface) == 0);
*wt_connp = &conn->iface;
/*
* Destroying the connection on error will destroy our session handle,
* cleanup using the session handle first, then discard the connection.
*/
err: if (cbuf != NULL)
__wt_buf_free(session, cbuf);
__wt_buf_free(session, &expath);
__wt_buf_free(session, &exconfig);
if (ret != 0 && conn != NULL)
WT_TRET(__wt_connection_close(conn));
/* Let the server threads proceed. */
if (ret == 0)
conn->connection_initialized = 1;
return (ret);
}
/*
* __wt_curjoin_join --
* Add a new join to a join cursor.
*/
int
__wt_curjoin_join(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_INDEX *idx, WT_CURSOR *ref_cursor, uint8_t flags, uint8_t range,
uint64_t count, uint32_t bloom_bit_count, uint32_t bloom_hash_count)
{
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_CURSOR_JOIN_ENDPOINT *end, *newend;
bool hasins, needbloom, range_eq;
u_int i, ins, nonbloom;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
char *main_uri;
size_t namesize, newsize;
entry = NULL;
hasins = needbloom = false;
ins = 0; /* -Wuninitialized */
main_uri = NULL;
nonbloom = 0; /* -Wuninitialized */
namesize = strlen(cjoin->table->name);
for (i = 0; i < cjoin->entries_next; i++) {
if (cjoin->entries[i].index == idx) {
entry = &cjoin->entries[i];
break;
}
if (!needbloom && i > 0 &&
!F_ISSET(&cjoin->entries[i], WT_CURJOIN_ENTRY_BLOOM)) {
needbloom = true;
nonbloom = i;
}
}
if (entry == NULL) {
WT_ERR(__wt_realloc_def(session, &cjoin->entries_allocated,
cjoin->entries_next + 1, &cjoin->entries));
if (LF_ISSET(WT_CURJOIN_ENTRY_BLOOM) && needbloom) {
/*
* Reorder the list so that after the first entry,
* the Bloom filtered entries come next, followed by
* the non-Bloom entries. Once the Bloom filters
* are built, determining membership via Bloom is
* faster than without Bloom, so we can answer
* membership questions more quickly, and with less
* I/O, with the Bloom entries first.
*/
entry = &cjoin->entries[nonbloom];
memmove(entry + 1, entry,
(cjoin->entries_next - nonbloom) *
sizeof(WT_CURSOR_JOIN_ENTRY));
memset(entry, 0, sizeof(WT_CURSOR_JOIN_ENTRY));
}
else
entry = &cjoin->entries[cjoin->entries_next];
entry->index = idx;
entry->flags = flags;
entry->count = count;
entry->bloom_bit_count = bloom_bit_count;
entry->bloom_hash_count = bloom_hash_count;
++cjoin->entries_next;
} else {
/* Merge the join into an existing entry for this index */
if (count != 0 && entry->count != 0 && entry->count != count)
WT_ERR_MSG(session, EINVAL,
"count=%" PRIu64 " does not match "
"previous count=%" PRIu64 " for this index",
count, entry->count);
if (LF_MASK(WT_CURJOIN_ENTRY_BLOOM) !=
F_MASK(entry, WT_CURJOIN_ENTRY_BLOOM))
WT_ERR_MSG(session, EINVAL,
"join has incompatible strategy "
"values for the same index");
/*
* Check against other comparisons (we call them endpoints)
* already set up for this index.
* We allow either:
* - one or more "eq" (with disjunction)
* - exactly one "eq" (with conjunction)
* - exactly one of "gt" or "ge" (conjunction or disjunction)
* - exactly one of "lt" or "le" (conjunction or disjunction)
* - one of "gt"/"ge" along with one of "lt"/"le"
* (currently restricted to conjunction).
*
* Some other combinations, although expressible either do
* not make sense (X == 3 AND X == 5) or are reducible (X <
* 7 AND X < 9). Other specific cases of (X < 7 OR X > 15)
* or (X == 4 OR X > 15) make sense but we don't handle yet.
*/
for (i = 0; i < entry->ends_next; i++) {
end = &entry->ends[i];
range_eq = (range == WT_CURJOIN_END_EQ);
if ((F_ISSET(end, WT_CURJOIN_END_GT) &&
((range & WT_CURJOIN_END_GT) != 0 || range_eq)) ||
(F_ISSET(end, WT_CURJOIN_END_LT) &&
((range & WT_CURJOIN_END_LT) != 0 || range_eq)) ||
(end->flags == WT_CURJOIN_END_EQ &&
(range & (WT_CURJOIN_END_LT | WT_CURJOIN_END_GT))
!= 0))
WT_ERR_MSG(session, EINVAL,
"join has overlapping ranges");
if (range == WT_CURJOIN_END_EQ &&
end->flags == WT_CURJOIN_END_EQ &&
!F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
WT_ERR_MSG(session, EINVAL,
"compare=eq can only be combined "
"using operation=or");
/*
* Sort "gt"/"ge" to the front, followed by any number
* of "eq", and finally "lt"/"le".
*/
if (!hasins &&
((range & WT_CURJOIN_END_GT) != 0 ||
(range == WT_CURJOIN_END_EQ &&
!F_ISSET(end, WT_CURJOIN_END_GT)))) {
ins = i;
hasins = true;
}
}
/* All checks completed, merge any new configuration now */
entry->count = count;
entry->bloom_bit_count =
WT_MAX(entry->bloom_bit_count, bloom_bit_count);
entry->bloom_hash_count =
WT_MAX(entry->bloom_hash_count, bloom_hash_count);
}
WT_ERR(__wt_realloc_def(session, &entry->ends_allocated,
entry->ends_next + 1, &entry->ends));
if (!hasins)
ins = entry->ends_next;
newend = &entry->ends[ins];
memmove(newend + 1, newend,
(entry->ends_next - ins) * sizeof(WT_CURSOR_JOIN_ENDPOINT));
memset(newend, 0, sizeof(WT_CURSOR_JOIN_ENDPOINT));
entry->ends_next++;
newend->cursor = ref_cursor;
F_SET(newend, range);
/* Open the main file with a projection of the indexed columns. */
if (entry->main == NULL && entry->index != NULL) {
namesize = strlen(cjoin->table->name);
newsize = namesize + entry->index->colconf.len + 1;
WT_ERR(__wt_calloc(session, 1, newsize, &main_uri));
snprintf(main_uri, newsize, "%s%.*s",
cjoin->table->name, (int)entry->index->colconf.len,
entry->index->colconf.str);
WT_ERR(__wt_open_cursor(session, main_uri,
(WT_CURSOR *)cjoin, raw_cfg, &entry->main));
}
err: if (main_uri != NULL)
__wt_free(session, main_uri);
return (ret);
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, int update, u_int start_chunk, uint32_t start_id)
{
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
int locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
lsm_tree = clsm->lsm_tree;
chunk = NULL;
ckpt_cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/* Copy the key, so we don't lose the cursor position. */
if (F_ISSET(c, WT_CURSTD_KEY_INT) && !WT_DATA_IN_ITEM(&c->key))
WT_RET(__wt_buf_set(
session, &c->key, c->key.data, c->key.size));
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
if (update) {
if (txn->isolation == TXN_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
} else
F_SET(clsm, WT_CLSM_OPEN_READ);
WT_RET(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
/*
* If there is no in-memory chunk in the tree for an update operation,
* create one.
*
* !!!
* It is exceeding unlikely that we get here at all, but if there is a
* transaction in progress and it rolls back, it would leave the
* metadata inconsistent.
*/
if (update && (lsm_tree->nchunks == 0 ||
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) == NULL ||
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))) {
/* Release our lock because switch will get a write lock. */
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
}
F_SET(session, WT_SESSION_NO_CACHE_CHECK);
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id) {
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->txnid_max));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->txnid_max[ngood - 1] =
chunk->txnid_max;
if (__wt_txn_visible_all(
session, chunk->txnid_max))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (clsm->cursors != NULL && (ngood < clsm->nchunks ||
(!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0))) {
saved_gen = lsm_tree->dsk_gen;
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0)
WT_ERR(__clsm_close_cursors(
clsm, 0, nchunks - nupdates));
WT_ERR(__clsm_close_cursors(
clsm, ngood, clsm->nchunks));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->txnid_max[i] = chunk->txnid_max;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL && !F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
WT_WITH_BTREE(session, ((WT_CURSOR_BTREE *)(primary))->btree,
__wt_btree_evictable(session, 0));
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err: F_CLR(session, WT_SESSION_NO_CACHE_CHECK);
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
return (ret);
}
/*
* __create_table --
* Create a table.
*/
static int
__create_table(WT_SESSION_IMPL *session,
const char *name, bool exclusive, const char *config)
{
WT_CONFIG conf;
WT_CONFIG_ITEM cgkey, cgval, cval;
WT_DECL_RET;
WT_TABLE *table;
const char *cfg[4] =
{ WT_CONFIG_BASE(session, table_meta), config, NULL, NULL };
const char *tablename;
char *tableconf, *cgname;
size_t cgsize;
int ncolgroups;
bool exists;
cgname = NULL;
table = NULL;
tableconf = NULL;
exists = false;
tablename = name;
if (!WT_PREFIX_SKIP(tablename, "table:"))
return (__wt_unexpected_object_type(session, name, "table:"));
if ((ret = __wt_schema_get_table(session,
tablename, strlen(tablename), false, &table)) == 0) {
if (exclusive)
WT_ERR(EEXIST);
exists = true;
}
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_gets(session, cfg, "colgroups", &cval));
__wt_config_subinit(session, &conf, &cval);
for (ncolgroups = 0;
(ret = __wt_config_next(&conf, &cgkey, &cgval)) == 0;
ncolgroups++)
;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_config_collapse(session, cfg, &tableconf));
if (!exists) {
WT_ERR(__wt_metadata_insert(session, name, tableconf));
/* Attempt to open the table now to catch any errors. */
WT_ERR(__wt_schema_get_table(
session, tablename, strlen(tablename), true, &table));
if (ncolgroups == 0) {
cgsize = strlen("colgroup:") + strlen(tablename) + 1;
WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
snprintf(cgname, cgsize, "colgroup:%s", tablename);
WT_ERR(__create_colgroup(
session, cgname, exclusive, config));
}
}
if (0) {
err: if (table != NULL) {
WT_TRET(__wt_schema_remove_table(session, table));
table = NULL;
}
}
if (table != NULL)
__wt_schema_release_table(session, table);
__wt_free(session, cgname);
__wt_free(session, tableconf);
return (ret);
}
