/*
* __curlog_close --
* WT_CURSOR.close method for the log cursor type.
*/
static int
__curlog_close(WT_CURSOR *cursor)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cl = (WT_CURSOR_LOG *)cursor;
CURSOR_API_CALL_PREPARE_ALLOWED(cursor, session, close, NULL);
err:
conn = S2C(session);
if (F_ISSET(cl, WT_CURLOG_ARCHIVE_LOCK)) {
(void)__wt_atomic_sub32(&conn->log_cursors, 1);
__wt_readunlock(session, &conn->log->log_archive_lock);
}
__wt_free(session, cl->cur_lsn);
__wt_free(session, cl->next_lsn);
__wt_scr_free(session, &cl->logrec);
__wt_scr_free(session, &cl->opkey);
__wt_scr_free(session, &cl->opvalue);
__wt_free(session, cl->packed_key);
__wt_free(session, cl->packed_value);
__wt_cursor_close(cursor);
API_END_RET(session, ret);
}
/*
* __lsm_unpin_chunks --
* Decrement the reference count for a set of chunks. Allowing those
* chunks to be considered for deletion.
*/
static void
__lsm_unpin_chunks(WT_SESSION_IMPL *session, WT_LSM_WORKER_COOKIE *cookie)
{
u_int i;
for (i = 0; i < cookie->nchunks; i++) {
if (cookie->chunk_array[i] == NULL)
continue;
WT_ASSERT(session, cookie->chunk_array[i]->refcnt > 0);
(void)__wt_atomic_sub32(&cookie->chunk_array[i]->refcnt, 1);
}
/* Ensure subsequent calls don't double decrement. */
cookie->nchunks = 0;
}
/*
* __lsm_worker_general_op --
* Execute a single bloom, drop or flush work unit.
*/
static int
__lsm_worker_general_op(
WT_SESSION_IMPL *session, WT_LSM_WORKER_ARGS *cookie, bool *completed)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORK_UNIT *entry;
bool force;
*completed = false;
/*
* Return if this thread cannot process a bloom, drop or flush.
*/
if (!FLD_ISSET(cookie->type,
WT_LSM_WORK_BLOOM | WT_LSM_WORK_DROP | WT_LSM_WORK_FLUSH))
return (WT_NOTFOUND);
if ((ret = __wt_lsm_manager_pop_entry(session,
cookie->type, &entry)) != 0 || entry == NULL)
return (ret);
if (entry->type == WT_LSM_WORK_FLUSH) {
force = F_ISSET(entry, WT_LSM_WORK_FORCE);
F_CLR(entry, WT_LSM_WORK_FORCE);
WT_ERR(__wt_lsm_get_chunk_to_flush(session,
entry->lsm_tree, force, &chunk));
/*
* If we got a chunk to flush, checkpoint it.
*/
if (chunk != NULL) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Flush%s chunk %d %s",
force ? " w/ force" : "",
chunk->id, chunk->uri));
ret = __wt_lsm_checkpoint_chunk(
session, entry->lsm_tree, chunk);
WT_ASSERT(session, chunk->refcnt > 0);
(void)__wt_atomic_sub32(&chunk->refcnt, 1);
WT_ERR(ret);
}
} else if (entry->type == WT_LSM_WORK_DROP)
WT_ERR(__wt_lsm_free_chunks(session, entry->lsm_tree));
else if (entry->type == WT_LSM_WORK_BLOOM)
WT_ERR(__wt_lsm_work_bloom(session, entry->lsm_tree));
*completed = true;
err: __wt_lsm_manager_free_work_unit(session, entry);
return (ret);
}
/*
* __wt_close --
* Close a file handle.
*/
int
__wt_close(WT_SESSION_IMPL *session, WT_FH **fhp)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *fh;
uint64_t bucket;
conn = S2C(session);
if (*fhp == NULL)
return (0);
fh = *fhp;
*fhp = NULL;
__wt_spin_lock(session, &conn->fh_lock);
if (fh == NULL || fh->ref == 0 || --fh->ref > 0) {
__wt_spin_unlock(session, &conn->fh_lock);
return (0);
}
/* Remove from the list. */
bucket = fh->name_hash % WT_HASH_ARRAY_SIZE;
WT_CONN_FILE_REMOVE(conn, fh, bucket);
(void)__wt_atomic_sub32(&conn->open_file_count, 1);
__wt_spin_unlock(session, &conn->fh_lock);
/* Discard the memory.
* Note: For directories, we do not open valid directory handles on
* windows since it is not possible to sync a directory
*/
if (fh->filehandle != INVALID_HANDLE_VALUE &&
CloseHandle(fh->filehandle) == 0) {
ret = __wt_errno();
__wt_err(session, ret, "CloseHandle: %s", fh->name);
}
if (fh->filehandle_secondary != INVALID_HANDLE_VALUE &&
CloseHandle(fh->filehandle_secondary) == 0) {
ret = __wt_errno();
__wt_err(session, ret, "CloseHandle: secondary: %s", fh->name);
}
__wt_free(session, fh->name);
__wt_free(session, fh);
return (ret);
}
/*
* __wt_close --
* Close a file handle.
*/
int
__wt_close(WT_SESSION_IMPL *session, WT_FH **fhp)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *fh;
uint64_t bucket;
conn = S2C(session);
if (*fhp == NULL)
return (0);
fh = *fhp;
*fhp = NULL;
/* Track handle-close as a file operation, so open and close match. */
WT_RET(__wt_verbose(
session, WT_VERB_FILEOPS, "%s: handle-close", fh->name));
/*
* If the reference count hasn't gone to 0, or if it's an in-memory
* object, we're done.
*
* Assert the reference count is correct, but don't let it wrap.
*/
__wt_spin_lock(session, &conn->fh_lock);
WT_ASSERT(session, fh->ref > 0);
if ((fh->ref > 0 && --fh->ref > 0) || F_ISSET(fh, WT_FH_IN_MEMORY)) {
__wt_spin_unlock(session, &conn->fh_lock);
return (0);
}
/* Remove from the list. */
bucket = fh->name_hash % WT_HASH_ARRAY_SIZE;
WT_CONN_FILE_REMOVE(conn, fh, bucket);
(void)__wt_atomic_sub32(&conn->open_file_count, 1);
__wt_spin_unlock(session, &conn->fh_lock);
/* Discard underlying resources. */
ret = fh->fh_close(session, fh);
__wt_free(session, fh->name);
__wt_free(session, fh);
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);
}