/*
* __wt_lsm_tree_truncate --
* Truncate an LSM tree.
*/
int
__wt_lsm_tree_truncate(
WT_SESSION_IMPL *session, const char *name, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
int locked;
WT_UNUSED(cfg);
chunk = NULL;
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, name, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_RET(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_RET(__wt_lsm_tree_lock(session, lsm_tree, 1));
locked = 1;
/* Create the new chunk. */
WT_ERR(__wt_calloc_def(session, 1, &chunk));
chunk->id = WT_ATOMIC_ADD(lsm_tree->last, 1);
WT_ERR(__wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
/* Mark all chunks old. */
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, 0, lsm_tree->nchunks, chunk));
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
WT_ERR(__lsm_tree_start_worker(session, lsm_tree));
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
__wt_lsm_tree_release(session, lsm_tree);
err: if (locked)
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
if (ret != 0) {
if (chunk != NULL) {
(void)__wt_schema_drop(session, chunk->uri, NULL);
__wt_free(session, chunk);
}
/*
* Discard the LSM tree structure on error. This will force the
* LSM tree to be re-opened the next time it is accessed and
* the last good version of the metadata will be used, resulting
* in a valid (not truncated) tree.
*/
WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
return (ret);
}
/*
* __wt_lsm_tree_drop --
* Drop an LSM tree.
*/
int
__wt_lsm_tree_drop(
WT_SESSION_IMPL *session, const char *name, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
u_int i;
int locked;
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, name, 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;
/* Drop the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
WT_ERR(__wt_schema_drop(session, chunk->uri, cfg));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(
__wt_schema_drop(session, chunk->bloom_uri, cfg));
}
/* Drop any chunks on the obsolete list. */
for (i = 0; i < lsm_tree->nold_chunks; i++) {
if ((chunk = lsm_tree->old_chunks[i]) == NULL)
continue;
WT_ERR(__wt_schema_drop(session, chunk->uri, cfg));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(
__wt_schema_drop(session, chunk->bloom_uri, cfg));
}
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
ret = __wt_metadata_remove(session, name);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __wt_lsm_tree_truncate --
* Truncate an LSM tree.
*/
int
__wt_lsm_tree_truncate(
WT_SESSION_IMPL *session, const char *name, const char *cfg[])
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
int locked;
WT_UNUSED(cfg);
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, name, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_RET(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_RET(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Create the new chunk. */
WT_ERR(__wt_calloc_def(session, 1, &chunk));
chunk->id = WT_ATOMIC_ADD(lsm_tree->last, 1);
WT_ERR(__wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
/* Mark all chunks old. */
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, 0, lsm_tree->nchunks, chunk));
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
WT_ERR(__lsm_tree_start_worker(session, lsm_tree));
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
/*
* Don't discard the LSM tree structure unless there has been an
* error. The handle remains valid for future operations.
*/
if (ret != 0)
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __wt_lsm_compact --
* Compact an LSM tree called via __wt_schema_worker.
*/
int
__wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, int *skip)
{
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
uint64_t last_merge_progressing;
time_t begin, end;
/*
* This function is applied to all matching sources: ignore anything
* that is not an LSM tree.
*/
if (!WT_PREFIX_MATCH(name, "lsm:"))
return (0);
/* Tell __wt_schema_worker not to look inside the LSM tree. */
*skip = 1;
WT_RET(__wt_lsm_tree_get(session, name, 0, &lsm_tree));
if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE) ||
lsm_tree->merge_threads == 0)
WT_RET_MSG(session, EINVAL,
"LSM compaction requires active merge threads");
WT_RET(__wt_seconds(session, &begin));
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
/* Wake up the merge threads. */
WT_RET(__wt_cond_signal(session, lsm_tree->work_cond));
/* Now wait for merge activity to stop. */
do {
last_merge_progressing = lsm_tree->merge_progressing;
__wt_sleep(1, 0);
WT_RET(__wt_seconds(session, &end));
if (session->compact->max_time > 0 &&
session->compact->max_time < (uint64_t)(end - begin))
WT_ERR(ETIMEDOUT);
} while (lsm_tree->merge_progressing != last_merge_progressing &&
lsm_tree->nchunks > 1);
err: F_CLR(lsm_tree, WT_LSM_TREE_COMPACTING);
return (ret);
}
/*
* __wt_lsm_tree_worker --
* Run a schema worker operation on each level of a LSM tree.
*/
int
__wt_lsm_tree_worker(WT_SESSION_IMPL *session,
const char *uri,
int (*file_func)(WT_SESSION_IMPL *, const char *[]),
int (*name_func)(WT_SESSION_IMPL *, const char *, int *),
const char *cfg[], uint32_t open_flags)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
u_int i;
int exclusive, locked;
locked = 0;
exclusive = FLD_ISSET(open_flags, WT_DHANDLE_EXCLUSIVE) ? 1 : 0;
WT_RET(__wt_lsm_tree_get(session, uri, exclusive, &lsm_tree));
/*
* We mark that we're busy using the tree to coordinate
* with merges so that merging doesn't change the chunk
* array out from underneath us.
*/
WT_ERR(exclusive ?
__wt_lsm_tree_writelock(session, lsm_tree) :
__wt_lsm_tree_readlock(session, lsm_tree));
locked = 1;
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
if (file_func == __wt_checkpoint &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))
continue;
WT_ERR(__wt_schema_worker(session, chunk->uri,
file_func, name_func, cfg, open_flags));
if (name_func == __wt_backup_list_uri_append &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(__wt_schema_worker(session, chunk->bloom_uri,
file_func, name_func, cfg, open_flags));
}
err: if (locked)
WT_TRET(exclusive ?
__wt_lsm_tree_writeunlock(session, lsm_tree) :
__wt_lsm_tree_readunlock(session, lsm_tree));
__wt_lsm_tree_release(session, lsm_tree);
return (ret);
}
/*
* __wt_lsm_tree_worker --
* Run a schema worker operation on each level of a LSM tree.
*/
int
__wt_lsm_tree_worker(WT_SESSION_IMPL *session,
const char *uri,
int (*func)(WT_SESSION_IMPL *, const char *[]),
const char *cfg[], uint32_t open_flags)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
u_int i;
WT_RET(__wt_lsm_tree_get(session, uri,
FLD_ISSET(open_flags, WT_BTREE_EXCLUSIVE) ? 1 : 0, &lsm_tree));
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
if (func == __wt_checkpoint &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))
continue;
WT_ERR(__wt_schema_worker(
session, chunk->uri, func, cfg, open_flags));
}
err: __wt_lsm_tree_release(session, lsm_tree);
return (ret);
}
/*
* __wt_lsm_tree_rename --
* Rename an LSM tree.
*/
int
__wt_lsm_tree_rename(WT_SESSION_IMPL *session,
const char *oldname, const char *newname, const char *cfg[])
{
WT_DECL_RET;
WT_ITEM buf;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
const char *old;
u_int i;
int locked;
old = NULL;
WT_CLEAR(buf);
locked = 0;
/* Get the LSM tree. */
WT_RET(__wt_lsm_tree_get(session, oldname, 1, &lsm_tree));
/* Shut down the LSM worker. */
WT_ERR(__lsm_tree_close(session, lsm_tree));
/* Prevent any new opens. */
WT_ERR(__wt_try_writelock(session, lsm_tree->rwlock));
locked = 1;
/* Set the new name. */
WT_ERR(__lsm_tree_set_name(session, lsm_tree, newname));
/* Rename the chunks. */
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
old = chunk->uri;
chunk->uri = NULL;
WT_ERR(__wt_lsm_tree_chunk_name(
session, lsm_tree, chunk->id, &buf));
chunk->uri = __wt_buf_steal(session, &buf, NULL);
WT_ERR(__wt_schema_rename(session, old, chunk->uri, cfg));
__wt_free(session, old);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
old = chunk->bloom_uri;
chunk->bloom_uri = NULL;
WT_ERR(__wt_lsm_tree_bloom_name(
session, lsm_tree, chunk->id, &buf));
chunk->bloom_uri = __wt_buf_steal(session, &buf, NULL);
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
WT_ERR(__wt_schema_rename(
session, old, chunk->uri, cfg));
__wt_free(session, old);
}
}
ret = __wt_rwunlock(session, lsm_tree->rwlock);
locked = 0;
if (ret == 0)
ret = __wt_lsm_meta_write(session, lsm_tree);
if (ret == 0)
ret = __wt_metadata_remove(session, oldname);
err: if (locked)
WT_TRET(__wt_rwunlock(session, lsm_tree->rwlock));
if (old != NULL)
__wt_free(session, old);
/*
* Discard this LSM tree structure. The first operation on the renamed
* tree will create a new one.
*/
WT_TRET(__lsm_tree_discard(session, lsm_tree));
return (ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] = API_CONF_DEFAULTS(session, create, config);
const char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_read(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_RET(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_newest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST);
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_NEWEST) ||
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST)))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_config", &cval));
if (cval.type == ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_chunk_size", &cval));
lsm_tree->chunk_size = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm_merge_threads", &cval));
lsm_tree->merge_threads = (uint32_t)cval.val;
/* Sanity check that api_data.py is in sync with lsm.h */
WT_ASSERT(session, lsm_tree->merge_threads <= WT_LSM_MAX_WORKERS);
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u", config));
lsm_tree->file_config = __wt_buf_steal(session, buf, NULL);
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __wt_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);
}
/*
* __wt_lsm_compact --
* Compact an LSM tree called via __wt_schema_worker.
*/
int
__wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, int *skip)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
time_t begin, end;
uint64_t progress;
int i, compacting, flushing, locked, ref;
compacting = flushing = locked = ref = 0;
chunk = NULL;
/*
* This function is applied to all matching sources: ignore anything
* that is not an LSM tree.
*/
if (!WT_PREFIX_MATCH(name, "lsm:"))
return (0);
/* Tell __wt_schema_worker not to look inside the LSM tree. */
*skip = 1;
WT_RET(__wt_lsm_tree_get(session, name, 0, &lsm_tree));
if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE))
WT_ERR_MSG(session, EINVAL,
"LSM compaction requires active merge threads");
WT_ERR(__wt_seconds(session, &begin));
/*
* Compacting has two distinct phases.
* 1. All in-memory chunks up to and including the current
* current chunk must be flushed. Normally, the flush code
* does not flush the last, in-use chunk, so we set a force
* flag to include that last chunk. We monitor the state of the
* last chunk and periodically push another forced flush work
* unit until it is complete.
* 2. After all flushing is done, we move onto the merging
* phase for compaction. Again, we monitor the state and
* continue to push merge work units until all merging is done.
*/
/* Lock the tree: single-thread compaction. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = 1;
/* Clear any merge throttle: compact throws out that calculation. */
lsm_tree->merge_throttle = 0;
lsm_tree->merge_aggressiveness = 0;
progress = lsm_tree->merge_progressing;
/* If another thread started a compact on this tree, we're done. */
if (F_ISSET(lsm_tree, WT_LSM_TREE_COMPACTING))
goto err;
/*
* Set the switch transaction on the current chunk, if it
* hasn't been set before. This prevents further writes, so it
* can be flushed by the checkpoint worker.
*/
if (lsm_tree->nchunks > 0 &&
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) != NULL) {
if (chunk->switch_txn == WT_TXN_NONE)
chunk->switch_txn = __wt_txn_new_id(session);
/*
* If we have a chunk, we want to look for it to be on-disk.
* So we need to add a reference to keep it available.
*/
(void)WT_ATOMIC_ADD4(chunk->refcnt, 1);
ref = 1;
}
locked = 0;
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (chunk != NULL) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Compact force flush %s flags 0x%" PRIx32
" chunk %u flags 0x%"
PRIx32, name, lsm_tree->flags, chunk->id, chunk->flags));
flushing = 1;
/*
* Make sure the in-memory chunk gets flushed do not push a
* switch, because we don't want to create a new in-memory
* chunk if the tree is being used read-only now.
*/
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_FLUSH, WT_LSM_WORK_FORCE, lsm_tree));
} else {
/*
* If there is no chunk to flush, go straight to the
* compacting state.
*/
compacting = 1;
progress = lsm_tree->merge_progressing;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"COMPACT: Start compacting %s", lsm_tree->name));
}
/* Wait for the work unit queues to drain. */
while (F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE)) {
/*
* The flush flag is cleared when the chunk has been flushed.
* Continue to push forced flushes until the chunk is on disk.
* Once it is on disk move to the compacting phase.
*/
if (flushing) {
WT_ASSERT(session, chunk != NULL);
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
WT_ERR(__wt_verbose(session,
WT_VERB_LSM,
"Compact flush done %s chunk %u. "
"Start compacting progress %" PRIu64,
name, chunk->id,
lsm_tree->merge_progressing));
(void)WT_ATOMIC_SUB4(chunk->refcnt, 1);
flushing = ref = 0;
compacting = 1;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
progress = lsm_tree->merge_progressing;
} else {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Compact flush retry %s chunk %u",
name, chunk->id));
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_FLUSH, WT_LSM_WORK_FORCE,
lsm_tree));
}
}
/*
* The compacting flag is cleared when no merges can be done.
* Ensure that we push through some aggressive merges before
* stopping otherwise we might not do merges that would
* span chunks with different generations.
*/
if (compacting && !F_ISSET(lsm_tree, WT_LSM_TREE_COMPACTING)) {
if (lsm_tree->merge_aggressiveness < 10 ||
(progress < lsm_tree->merge_progressing) ||
lsm_tree->merge_syncing) {
progress = lsm_tree->merge_progressing;
F_SET(lsm_tree, WT_LSM_TREE_COMPACTING);
lsm_tree->merge_aggressiveness = 10;
} else
break;
}
__wt_sleep(1, 0);
WT_ERR(__wt_seconds(session, &end));
if (session->compact->max_time > 0 &&
session->compact->max_time < (uint64_t)(end - begin)) {
WT_ERR(ETIMEDOUT);
}
/*
* Push merge operations while they are still getting work
* done. If we are pushing merges, make sure they are
* aggressive, to avoid duplicating effort.
*/
if (compacting)
#define COMPACT_PARALLEL_MERGES 5
for (i = lsm_tree->queue_ref;
i < COMPACT_PARALLEL_MERGES; i++) {
lsm_tree->merge_aggressiveness = 10;
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
}
}
err:
/* Ensure anything we set is cleared. */
if (ref)
(void)WT_ATOMIC_SUB4(chunk->refcnt, 1);
if (compacting) {
F_CLR(lsm_tree, WT_LSM_TREE_COMPACTING);
lsm_tree->merge_aggressiveness = 0;
}
if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Compact %s complete, return %d", name, ret));
__wt_lsm_tree_release(session, lsm_tree);
return (ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
const char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_search(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "collator", &cval));
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->collator_name));
WT_ERR(__wt_config_gets(session, cfg, "lsm.auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(session, cval.str, cval.len,
&lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_max", &cval));
lsm_tree->chunk_max = (uint64_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_size", &cval));
lsm_tree->chunk_size = (uint64_t)cval.val;
if (lsm_tree->chunk_size > lsm_tree->chunk_max)
WT_ERR_MSG(session, EINVAL,
"Chunk size (chunk_size) must be smaller than or equal to "
"the maximum chunk size (chunk_max)");
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
lsm_tree->merge_min = lsm_tree->merge_max / 2;
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_threads", &cval));
lsm_tree->merge_threads = (uint32_t)cval.val;
/* Sanity check that api_data.py is in sync with lsm.h */
WT_ASSERT(session, lsm_tree->merge_threads <= WT_LSM_MAX_WORKERS);
/*
* Set up the config for each chunk. If possible, avoid high latencies
* from fsync by flushing the cache every 8MB (will be overridden by
* any application setting).
*/
tmpconfig = "";
#ifdef HAVE_SYNC_FILE_RANGE
if (!S2C(session)->direct_io)
tmpconfig = "os_cache_dirty_max=8MB,";
#endif
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s%s,key_format=u,value_format=u", tmpconfig, config));
lsm_tree->file_config = __wt_buf_steal(session, buf, NULL);
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}