/*
* __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;
u_int created_chunk, verb;
int create_bloom, locked, in_sync, tret;
const char *cfg[3];
const char *drop_cfg[] =
{ WT_CONFIG_BASE(session, session_drop), "force", NULL };
bloom = NULL;
chunk = NULL;
create_bloom = 0;
created_chunk = 0;
dest = src = NULL;
locked = 0;
start_id = 0;
in_sync = 0;
/* 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 = 1;
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 = 0;
/* Allocate an ID for the merge. */
dest_id = WT_ATOMIC_ADD4(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,
"%s: Chunk[%u] id %u",
lsm_tree->name, verb, lsm_tree->chunk[verb]->id));
}
WT_ERR(__wt_calloc_one(session, &chunk));
created_chunk = 1;
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 = 1;
/*
* 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, 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 1000
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_ADD4(lsm_tree->merge_syncing, 1);
in_sync = 1;
/*
* 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_CACHE_CHECK);
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_SUB4(lsm_tree->merge_syncing, 1);
in_sync = 0;
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 = 1;
/*
* 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, 1);
/* 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_SUB4(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", wiredtiger_strerror(ret)));
}
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_CACHE_CHECK);
return (ret);
}
/*
* __create_colgroup --
* Create a column group.
*/
static int
__create_colgroup(WT_SESSION_IMPL *session,
const char *name, bool 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, *tablename;
char *cgconf, *sourceconf, *oldconf;
cgconf = sourceconf = 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, true, &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, true, &fmt));
}
sourcecfg[1] = fmt.data;
WT_ERR(__wt_config_concat(session, sourcecfg, &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);
}
/*
* __wt_lsm_tree_rename --
* Rename an LSM tree.
*/
int
__wt_lsm_tree_rename(WT_SESSION_IMPL *session,
const char *olduri, const char *newuri, const char *cfg[])
{
WT_DECL_RET;
WT_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, olduri, 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, newuri));
/* Rename the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
old = chunk->uri;
chunk->uri = NULL;
WT_ERR(__wt_lsm_tree_chunk_name(
session, lsm_tree, chunk->id, &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, olduri);
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);
}
