/*
* __wt_lsm_work_switch --
* Do a switch if the LSM tree needs one.
*/
int
__wt_lsm_work_switch(
WT_SESSION_IMPL *session, WT_LSM_WORK_UNIT **entryp, bool *ran)
{
WT_DECL_RET;
WT_LSM_WORK_UNIT *entry;
/* We've become responsible for freeing the work unit. */
entry = *entryp;
*ran = false;
*entryp = NULL;
if (F_ISSET(entry->lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_lsm_tree_switch(session, entry->lsm_tree));
/* Failing to complete the switch is fine */
if (ret == EBUSY) {
if (F_ISSET(entry->lsm_tree, WT_LSM_TREE_NEED_SWITCH))
WT_ERR(__wt_lsm_manager_push_entry(session,
WT_LSM_WORK_SWITCH, 0, entry->lsm_tree));
ret = 0;
} else
*ran = true;
}
err: __wt_lsm_manager_free_work_unit(session, entry);
return (ret);
}
/*__wt_session_compact*/
int __wt_session_compact(WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_COMPACT compact;
WT_CONFIG_ITEM cval;
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, compact, config, cfg);
/* Setup the structure in the session handle */
memset(&compact, 0, sizeof(WT_COMPACT));
session->compact = &compact;
WT_ERR(__wt_config_gets(session, cfg, "timeout", &cval));
session->compact->max_time = (uint64_t)cval.val;
/* Find the types of data sources are being compacted. */
WT_WITH_SCHEMA_LOCK(session, ret = __wt_schema_worker(session, uri, NULL, __wt_compact_uri_analyze, cfg, 0));
WT_ERR(ret);
if (session->compact->lsm_count != 0)
WT_ERR(__wt_schema_worker(session, uri, NULL, __wt_lsm_compact, cfg, 0));
if (session->compact->file_count != 0)
WT_ERR(__compact_file(session, uri, cfg));
err:
session->compact = NULL;
API_END_RET_NOTFOUND_MAP(session, ret);
}
/*
* __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_RET(__lsm_discard_handle(session, uri, WT_CHECKPOINT));
/*
* 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 = __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_curbackup_open --
* WT_SESSION->open_cursor method for the backup cursor type.
*/
int
__wt_curbackup_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value_notsup, /* get-value */
__wt_cursor_set_key_notsup, /* set-key */
__wt_cursor_set_value_notsup, /* set-value */
__wt_cursor_compare_notsup, /* compare */
__wt_cursor_equals_notsup, /* equals */
__curbackup_next, /* next */
__wt_cursor_notsup, /* prev */
__curbackup_reset, /* reset */
__wt_cursor_notsup, /* search */
__wt_cursor_search_near_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_modify_notsup, /* modify */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_notsup, /* reserve */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curbackup_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_BACKUP *cb;
WT_DECL_RET;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_BACKUP, iface) == 0);
cb = NULL;
WT_RET(__wt_calloc_one(session, &cb));
cursor = &cb->iface;
*cursor = iface;
cursor->session = &session->iface;
session->bkp_cursor = cb;
cursor->key_format = "S"; /* Return the file names as the key. */
cursor->value_format = ""; /* No value. */
/*
* Start the backup and fill in the cursor's list. Acquire the schema
* lock, we need a consistent view when creating a copy.
*/
WT_WITH_CHECKPOINT_LOCK(session,
WT_WITH_SCHEMA_LOCK(session,
ret = __backup_start(session, cb, cfg)));
WT_ERR(ret);
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
if (0) {
err: WT_TRET(__curbackup_close(cursor));
*cursorp = NULL;
}
return (ret);
}
/*
* __wt_clsm_open_bulk --
* WT_SESSION->open_cursor method for LSM bulk cursors.
*/
int
__wt_clsm_open_bulk(WT_CURSOR_LSM *clsm, const char *cfg[])
{
WT_CURSOR *cursor, *bulk_cursor;
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
bulk_cursor = NULL;
cursor = &clsm->iface;
lsm_tree = clsm->lsm_tree;
session = (WT_SESSION_IMPL *)clsm->iface.session;
F_SET(clsm, WT_CLSM_BULK);
/* Bulk cursors are limited to insert and close. */
__wt_cursor_set_notsup(cursor);
cursor->insert = __clsm_insert_bulk;
cursor->close = __clsm_close_bulk;
/*
* Setup the first chunk in the tree. This is the only time we switch
* without using the LSM worker threads, it's safe to do here since
* we have an exclusive lock on the LSM tree. We need to do this
* switch inline, since switch needs a schema lock and online index
* creation opens a bulk cursor while holding the schema lock.
*/
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_lsm_tree_switch(session, lsm_tree));
WT_RET(ret);
/*
* Open a bulk cursor on the first chunk, it's not a regular LSM chunk
* cursor, but use the standard storage locations. Allocate the space
* for a bloom filter - it makes cleanup simpler. Cleaned up by
* cursor close on error.
*/
WT_RET(__wt_calloc_one(session, &clsm->blooms));
clsm->bloom_alloc = 1;
WT_RET(__wt_calloc_one(session, &clsm->cursors));
clsm->cursor_alloc = 1;
clsm->nchunks = 1;
/*
* Open a bulk cursor on the first chunk in the tree - take a read
* lock on the LSM tree while we are opening the chunk, to ensure
* that the first chunk has been fully created before we succeed.
* Pass through the application config to ensure the tree is open
* for bulk access.
*/
WT_RET(__wt_open_cursor(session,
lsm_tree->chunk[0]->uri, &clsm->iface, cfg, &bulk_cursor));
clsm->cursors[0] = bulk_cursor;
/* LSM cursors are always raw */
F_SET(bulk_cursor, WT_CURSTD_RAW);
return (0);
}
/*btree file的compact操作*/
static int __compact_file(WT_SESSION_IMPL* session, const char* uri, const char* cfg[])
{
WT_DECL_RET;
WT_DECL_ITEM(t);
WT_SESSION *wt_session;
WT_TXN *txn;
int i;
struct timespec start_time;
txn = &session->txn;
wt_session = &session->iface;
/*
* File compaction requires checkpoints, which will fail in a
* transactional context. Check now so the error message isn't
* confusing.
*/
if(session->compact->file_count != 0 && F_ISSET(txn, TXN_RUNNING))
WT_ERR_MSG(session, EINVAL, " File compaction not permitted in a transaction");
/*
* Force the checkpoint: we don't want to skip it because the work we
* need to have done is done in the underlying block manager.
*/
WT_ERR(__wt_scr_alloc(session, 128, &t));
WT_ERR(__wt_buf_fmt(session, t, "target=(\"%s\"),force=1", uri));
WT_ERR(__wt_epoch(session, &start_time));
/*
* We compact 10% of the file on each pass (but the overall size of the
* file is decreasing each time, so we're not compacting 10% of the
* original file each time). Try 100 times (which is clearly more than
* we need); quit if we make no progress and check for a timeout each
* time through the loop.
*/
for (i = 0; i < 100; ++i) {
WT_ERR(wt_session->checkpoint(wt_session, t->data));
session->compaction = 0;
WT_WITH_SCHEMA_LOCK(session, ret = __wt_schema_worker(session, uri, __wt_compact, NULL, cfg, 0));
WT_ERR(ret);
if (!session->compaction)
break;
WT_ERR(wt_session->checkpoint(wt_session, t->data));
WT_ERR(wt_session->checkpoint(wt_session, t->data));
WT_ERR(__session_compact_check_timeout(session, start_time));
}
err:
__wt_scr_free(session, &t);
}
/*
* __wt_curbackup_open --
* WT_SESSION->open_cursor method for the backup cursor type.
*/
int
__wt_curbackup_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
NULL, /* get-key */
__wt_cursor_notsup, /* get-value */
__wt_cursor_notsup, /* set-key */
__wt_cursor_notsup, /* set-value */
NULL, /* compare */
__curbackup_next, /* next */
__wt_cursor_notsup, /* prev */
__curbackup_reset, /* reset */
__wt_cursor_notsup, /* search */
__wt_cursor_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__curbackup_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_BACKUP *cb;
WT_DECL_RET;
STATIC_ASSERT(offsetof(WT_CURSOR_BACKUP, iface) == 0);
cb = NULL;
WT_RET(__wt_calloc_def(session, 1, &cb));
cursor = &cb->iface;
*cursor = iface;
cursor->session = &session->iface;
session->bkp_cursor = cb;
cursor->key_format = "S"; /* Return the file names as the key. */
cursor->value_format = ""; /* No value. */
/*
* Start the backup and fill in the cursor's list. Acquire the schema
* lock, we need a consistent view when creating a copy.
*/
WT_WITH_SCHEMA_LOCK(session, ret = __backup_start(session, cb, cfg));
WT_ERR(ret);
/* __wt_cursor_init is last so we don't have to clean up on error. */
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
if (0) {
err: __wt_free(session, cb);
}
return (ret);
}
/*
* __metadata_init --
* Create the metadata file.
*/
static int
__metadata_init(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
/*
* 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_METAFILE_URI, NULL));
return (ret);
}
/*
* __session_verify --
* WT_SESSION->verify method.
*/
static int
__session_verify(WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, verify, config, cfg);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, uri,
__wt_verify, cfg, WT_BTREE_EXCLUSIVE | WT_BTREE_VERIFY));
err: API_END_NOTFOUND_MAP(session, ret);
}
/*
* __session_drop --
* WT_SESSION->drop method.
*/
static int
__session_drop(WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, drop, config, cfg);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_drop(session, uri, cfg));
err: /* Note: drop operations cannot be unrolled (yet?). */
API_END_NOTFOUND_MAP(session, ret);
}
/*
* __session_compact_worker --
* Worker function to do the actual compaction call.
*/
static int
__session_compact_worker(
WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, compact, config, cfg);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, uri, __wt_compact, cfg, 0));
err: API_END_NOTFOUND_MAP(session, ret);
}
/*
* __session_rename --
* WT_SESSION->rename method.
*/
static int
__session_rename(WT_SESSION *wt_session,
const char *uri, const char *newname, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, rename, config, cfg);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_rename(session, uri, newname, cfg));
err: API_END_NOTFOUND_MAP(session, ret);
}
/*
* __session_create --
* WT_SESSION->create method.
*/
static int
__session_create(WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, create, config, cfg);
WT_UNUSED(cfg);
/* Disallow objects in the WiredTiger name space. */
WT_ERR(__wt_schema_name_check(session, uri));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_create(session, uri, config));
err: API_END_NOTFOUND_MAP(session, ret);
}
/*
* __session_checkpoint --
* WT_SESSION->checkpoint method.
*/
static int
__session_checkpoint(WT_SESSION *wt_session, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_TXN *txn;
session = (WT_SESSION_IMPL *)wt_session;
txn = &session->txn;
WT_CSTAT_INCR(session, checkpoint);
SESSION_API_CALL(session, checkpoint, config, cfg);
/*
* Checkpoints require a snapshot to write a transactionally consistent
* snapshot of the data.
*
* We can't use an application's transaction: if it has uncommitted
* changes, they will be written in the checkpoint and may appear after
* a crash.
*
* Use a real snapshot transaction: we don't want any chance of the
* snapshot being updated during the checkpoint. Eviction is prevented
* from evicting anything newer than this because we track the oldest
* transaction ID in the system that is not visible to all readers.
*/
if (F_ISSET(txn, TXN_RUNNING))
WT_ERR_MSG(session, EINVAL,
"Checkpoint not permitted in a transaction");
/*
* Reset open cursors.
*
* We do this here explicitly even though it will happen implicitly in
* the call to begin_transaction for the checkpoint, in case some
* implementation of WT_CURSOR::reset needs the schema lock.
*/
WT_ERR(__session_reset_cursors(session));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_txn_checkpoint(session, cfg));
err: API_END_NOTFOUND_MAP(session, 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_clsm_init_merge --
* Initialize an LSM cursor for a merge.
*/
int
__wt_clsm_init_merge(
WT_CURSOR *cursor, u_int start_chunk, uint32_t start_id, u_int nchunks)
{
WT_CURSOR_LSM *clsm;
WT_DECL_RET;
WT_SESSION_IMPL *session;
clsm = (WT_CURSOR_LSM *)cursor;
session = (WT_SESSION_IMPL *)cursor->session;
F_SET(clsm, WT_CLSM_MERGE);
if (start_chunk != 0)
F_SET(clsm, WT_CLSM_MINOR_MERGE);
clsm->nchunks = nchunks;
WT_WITH_SCHEMA_LOCK(session,
ret = __clsm_open_cursors(clsm, 0, start_chunk, start_id));
return (ret);
}
/*
* __curbackup_close --
* WT_CURSOR->close method for the backup cursor type.
*/
static int
__curbackup_close(WT_CURSOR *cursor)
{
WT_CURSOR_BACKUP *cb;
WT_DECL_RET;
WT_SESSION_IMPL *session;
int tret;
cb = (WT_CURSOR_BACKUP *)cursor;
CURSOR_API_CALL(cursor, session, close, NULL);
WT_TRET(__backup_cleanup_handles(session, cb));
WT_TRET(__wt_cursor_close(cursor));
session->bkp_cursor = NULL;
WT_WITH_SCHEMA_LOCK(session,
tret = __backup_stop(session)); /* Stop the backup. */
WT_TRET(tret);
err: API_END_RET(session, ret);
}
/*
* __wt_lsm_checkpoint_chunk --
* Flush a single LSM chunk to disk.
*/
int
__wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_DECL_RET;
WT_TXN_ISOLATION saved_isolation;
bool flush_set;
flush_set = false;
/*
* If the chunk is already checkpointed, make sure it is also evicted.
* Either way, there is no point trying to checkpoint it again.
*/
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!F_ISSET(chunk, WT_LSM_CHUNK_STABLE) &&
!chunk->evicted) {
WT_WITH_HANDLE_LIST_LOCK(session,
ret = __lsm_discard_handle(session, chunk->uri, NULL));
if (ret == 0)
chunk->evicted = 1;
else if (ret == EBUSY)
ret = 0;
else
WT_RET_MSG(session, ret, "discard handle");
}
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s already on disk",
chunk->uri));
return (0);
}
/* Stop if a running transaction needs the chunk. */
__wt_txn_update_oldest(session, true);
if (chunk->switch_txn == WT_TXN_NONE ||
!__wt_txn_visible_all(session, chunk->switch_txn)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s: running transaction, return",
chunk->uri));
return (0);
}
if (!__wt_atomic_cas8(&chunk->flushing, 0, 1))
return (0);
flush_set = true;
WT_ERR(__wt_verbose(session, WT_VERB_LSM, "LSM worker flushing %s",
chunk->uri));
/*
* Flush the file before checkpointing: this is the expensive part in
* terms of I/O.
*
* !!!
* We can wait here for checkpoints and fsyncs to complete, which can
* take a long time.
*/
if ((ret = __wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0)) == 0) {
/*
* Set read-uncommitted: we have already checked that all of the
* updates in this chunk are globally visible, use the cheapest
* possible check in reconciliation.
*/
saved_isolation = session->txn.isolation;
session->txn.isolation = WT_ISO_READ_UNCOMMITTED;
ret = __wt_cache_op(session, NULL, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
WT_TRET(__wt_session_release_btree(session));
}
WT_ERR(ret);
WT_ERR(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s",
chunk->uri));
/*
* Turn on metadata tracking to ensure the checkpoint gets the
* necessary handle locks.
*
* Ensure that we don't race with a running checkpoint: the checkpoint
* lock protects against us racing with an application checkpoint in
* this chunk. Don't wait for it, though: checkpoints can take a long
* time, and our checkpoint operation should be very quick.
*/
WT_ERR(__wt_meta_track_on(session));
WT_WITH_CHECKPOINT_LOCK(session, ret,
WT_WITH_SCHEMA_LOCK(session, ret,
ret = __wt_schema_worker(
session, chunk->uri, __wt_checkpoint, NULL, NULL, 0)));
WT_TRET(__wt_meta_track_off(session, false, ret != 0));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM checkpoint");
/* Now the file is written, get the chunk size. */
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
/* Update the flush timestamp to help track ongoing progress. */
WT_ERR(__wt_epoch(session, &lsm_tree->last_flush_ts));
++lsm_tree->chunks_flushed;
/* Lock the tree, mark the chunk as on disk and update the metadata. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
/* Update the throttle time. */
__wt_lsm_tree_throttle(session, lsm_tree, true);
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM metadata write");
WT_PUBLISH(chunk->flushing, 0);
flush_set = false;
/*
* Clear the no-eviction flag so the primary can be evicted and
* eventually closed. Only do this once the checkpoint has succeeded:
* otherwise, accessing the leaf page during the checkpoint can trigger
* forced eviction.
*/
WT_ERR(__wt_session_get_btree(session, chunk->uri, NULL, NULL, 0));
__wt_btree_evictable(session, true);
WT_ERR(__wt_session_release_btree(session));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
WT_ERR(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointed %s",
chunk->uri));
/* Schedule a bloom filter create for our newly flushed chunk. */
if (!FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF))
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_BLOOM, 0, lsm_tree));
else
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
err: if (flush_set)
WT_PUBLISH(chunk->flushing, 0);
return (ret);
}
/*
* __wt_lsm_checkpoint_chunk --
* Flush a single LSM chunk to disk.
*/
int
__wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk)
{
WT_DECL_RET;
WT_TXN_ISOLATION saved_isolation;
/*
* If the chunk is already checkpointed, make sure it is also evicted.
* Either way, there is no point trying to checkpoint it again.
*/
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!F_ISSET(chunk, WT_LSM_CHUNK_STABLE) &&
!chunk->evicted) {
if ((ret = __lsm_discard_handle(
session, chunk->uri, NULL)) == 0)
chunk->evicted = 1;
else if (ret == EBUSY)
ret = 0;
else
WT_RET_MSG(session, ret, "discard handle");
}
if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s already on disk",
chunk->uri));
return (0);
}
/* Stop if a running transaction needs the chunk. */
__wt_txn_update_oldest(session);
if (chunk->switch_txn == WT_TXN_NONE ||
!__wt_txn_visible_all(session, chunk->switch_txn)) {
WT_RET(__wt_verbose(session, WT_VERB_LSM,
"LSM worker %s: running transaction, return",
chunk->uri));
return (0);
}
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker flushing %s",
chunk->uri));
/*
* Flush the file before checkpointing: this is the expensive part in
* terms of I/O.
*
* Use the special eviction isolation level to avoid interfering with
* an application checkpoint: we have already checked that all of the
* updates in this chunk are globally visible.
*
* !!! We can wait here for checkpoints and fsyncs to complete, which
* can be a long time.
*/
if ((ret = __wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0)) == 0) {
saved_isolation = session->txn.isolation;
session->txn.isolation = TXN_ISO_EVICTION;
ret = __wt_cache_op(session, NULL, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
WT_TRET(__wt_session_release_btree(session));
}
WT_RET(ret);
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s",
chunk->uri));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, chunk->uri,
__wt_checkpoint, NULL, NULL, 0));
if (ret != 0)
WT_RET_MSG(session, ret, "LSM checkpoint");
/* Now the file is written, get the chunk size. */
WT_RET(__wt_lsm_tree_set_chunk_size(session, chunk));
/* Update the flush timestamp to help track ongoing progress. */
WT_RET(__wt_epoch(session, &lsm_tree->last_flush_ts));
/* Lock the tree, mark the chunk as on disk and update the metadata. */
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
/* Update the throttle time. */
__wt_lsm_tree_throttle(session, lsm_tree, 1);
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (ret != 0)
WT_RET_MSG(session, ret, "LSM metadata write");
/*
* Clear the no-eviction flag so the primary can be evicted and
* eventually closed. Only do this once the checkpoint has succeeded:
* otherwise, accessing the leaf page during the checkpoint can trigger
* forced eviction.
*/
WT_RET(__wt_session_get_btree(session, chunk->uri, NULL, NULL, 0));
__wt_btree_evictable(session, 1);
WT_RET(__wt_session_release_btree(session));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
WT_RET(__wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointed %s",
chunk->uri));
/* Schedule a bloom filter create for our newly flushed chunk. */
if (!FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF))
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_BLOOM, 0, lsm_tree));
else
WT_RET(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
return (0);
}
/*
* 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 WT_CONNECTION stdc = {
__conn_close,
__conn_reconfigure,
__conn_get_home,
__conn_is_new,
__conn_open_session,
__conn_load_extension,
__conn_add_data_source,
__conn_add_collator,
__conn_add_compressor,
__conn_add_extractor
};
static 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[] =
{ __wt_confdfl_wiredtiger_open, config, NULL, NULL, NULL };
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 the configuration strings. */
WT_ERR(__wt_config_check(
session, __wt_confchk_wiredtiger_open, config, 0));
/* 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));
/*
* XXX LSM initialization.
* This is structured so that it could be moved to an extension.
*/
WT_ERR(__wt_lsm_init(&conn->iface, NULL));
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_destroy(conn));
/* Let the server threads proceed. */
if (ret == 0)
conn->connection_initialized = 1;
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);
}
/*
* __clsm_enter --
* Start an operation on an LSM cursor, update if the tree has changed.
*/
static inline int
__clsm_enter(WT_CURSOR_LSM *clsm, bool reset, bool update)
{
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
uint64_t *switch_txnp;
uint64_t snap_min;
lsm_tree = clsm->lsm_tree;
session = (WT_SESSION_IMPL *)clsm->iface.session;
txn = &session->txn;
/* Merge cursors never update. */
if (F_ISSET(clsm, WT_CLSM_MERGE))
return (0);
if (reset) {
WT_ASSERT(session, !F_ISSET(&clsm->iface,
WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT));
WT_RET(__clsm_reset_cursors(clsm, NULL));
}
for (;;) {
/*
* If the cursor looks up-to-date, check if the cache is full.
* In case this call blocks, the check will be repeated before
* proceeding.
*/
if (clsm->dsk_gen != lsm_tree->dsk_gen &&
lsm_tree->nchunks != 0)
goto open;
if (clsm->dsk_gen != lsm_tree->dsk_gen &&
lsm_tree->nchunks != 0)
goto open;
/* Update the maximum transaction ID in the primary chunk. */
if (update) {
/*
* Ensure that there is a transaction snapshot active.
*/
WT_RET(__wt_txn_autocommit_check(session));
WT_RET(__wt_txn_id_check(session));
WT_RET(__clsm_enter_update(clsm));
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
if (txn->isolation == WT_ISO_SNAPSHOT)
__wt_txn_cursor_op(session);
/*
* Figure out how many updates are required for
* snapshot isolation.
*
* This is not a normal visibility check on the maximum
* transaction ID in each chunk: any transaction ID
* that overlaps with our snapshot is a potential
* conflict.
*/
clsm->nupdates = 1;
if (txn->isolation == WT_ISO_SNAPSHOT &&
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
WT_ASSERT(session,
F_ISSET(txn, WT_TXN_HAS_SNAPSHOT));
snap_min = txn->snap_min;
for (switch_txnp =
&clsm->switch_txn[clsm->nchunks - 2];
clsm->nupdates < clsm->nchunks;
clsm->nupdates++, switch_txnp--) {
if (WT_TXNID_LT(*switch_txnp, snap_min))
break;
WT_ASSERT(session,
!__wt_txn_visible_all(
session, *switch_txnp));
}
}
}
/*
* Stop when we are up-to-date, as long as this is:
* - a snapshot isolation update and the cursor is set up for
* that;
* - an update operation with a primary chunk, or
* - a read operation and the cursor is open for reading.
*/
if ((!update ||
txn->isolation != WT_ISO_SNAPSHOT ||
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) &&
((update && clsm->primary_chunk != NULL) ||
(!update && F_ISSET(clsm, WT_CLSM_OPEN_READ))))
break;
open: WT_WITH_SCHEMA_LOCK(session,
ret = __clsm_open_cursors(clsm, update, 0, 0));
WT_RET(ret);
}
if (!F_ISSET(clsm, WT_CLSM_ACTIVE)) {
WT_RET(__cursor_enter(session));
F_SET(clsm, WT_CLSM_ACTIVE);
}
return (0);
}
/*
* __wt_session_compact --
* WT_SESSION.compact method.
*/
int
__wt_session_compact(
WT_SESSION *wt_session, const char *uri, const char *config)
{
WT_COMPACT_STATE compact;
WT_CONFIG_ITEM cval;
WT_DATA_SOURCE *dsrc;
WT_DECL_RET;
WT_SESSION_IMPL *session;
u_int i;
bool ignore_cache_size_set;
ignore_cache_size_set = false;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_API_CALL(session, compact, config, cfg);
/*
* The compaction thread should not block when the cache is full: it is
* holding locks blocking checkpoints and once the cache is full, it can
* spend a long time doing eviction.
*/
if (!F_ISSET(session, WT_SESSION_IGNORE_CACHE_SIZE)) {
ignore_cache_size_set = true;
F_SET(session, WT_SESSION_IGNORE_CACHE_SIZE);
}
/* In-memory ignores compaction operations. */
if (F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto err;
/*
* Non-LSM object compaction requires checkpoints, which are impossible
* in transactional contexts. Disallow in all contexts (there's no
* reason for LSM to allow this, possible or not), and check now so the
* error message isn't confusing.
*/
WT_ERR(__wt_txn_context_check(session, false));
/* Disallow objects in the WiredTiger name space. */
WT_ERR(__wt_str_name_check(session, uri));
if (!WT_PREFIX_MATCH(uri, "colgroup:") &&
!WT_PREFIX_MATCH(uri, "file:") &&
!WT_PREFIX_MATCH(uri, "index:") &&
!WT_PREFIX_MATCH(uri, "lsm:") &&
!WT_PREFIX_MATCH(uri, "table:")) {
if ((dsrc = __wt_schema_get_source(session, uri)) != NULL)
ret = dsrc->compact == NULL ?
__wt_object_unsupported(session, uri) :
dsrc->compact(
dsrc, wt_session, uri, (WT_CONFIG_ARG *)cfg);
else
ret = __wt_bad_object_type(session, uri);
goto err;
}
/* Setup the session handle's compaction state structure. */
memset(&compact, 0, sizeof(WT_COMPACT_STATE));
session->compact = &compact;
/* Compaction can be time-limited. */
WT_ERR(__wt_config_gets(session, cfg, "timeout", &cval));
session->compact->max_time = (uint64_t)cval.val;
__wt_epoch(session, &session->compact->begin);
/* Find the types of data sources being compacted. */
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, uri,
__compact_handle_append, __compact_uri_analyze, cfg, 0));
WT_ERR(ret);
if (session->compact->lsm_count != 0)
WT_ERR(__wt_schema_worker(
session, uri, NULL, __wt_lsm_compact, cfg, 0));
if (session->compact->file_count != 0)
WT_ERR(__compact_worker(session));
err: session->compact = NULL;
for (i = 0; i < session->op_handle_next; ++i) {
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__compact_end(session)));
WT_WITH_DHANDLE(session, session->op_handle[i],
WT_TRET(__wt_session_release_dhandle(session)));
}
__wt_free(session, session->op_handle);
session->op_handle_allocated = session->op_handle_next = 0;
/*
* Release common session resources (for example, checkpoint may acquire
* significant reconciliation structures/memory).
*/
WT_TRET(__wt_session_release_resources(session));
if (ignore_cache_size_set)
F_CLR(session, WT_SESSION_IGNORE_CACHE_SIZE);
if (ret != 0)
WT_STAT_CONN_INCR(session, session_table_compact_fail);
else
WT_STAT_CONN_INCR(session, session_table_compact_success);
API_END_RET_NOTFOUND_MAP(session, ret);
}
/*
* __wt_lsm_checkpoint_worker --
* A worker thread for an LSM tree, responsible for flushing new chunks to
* disk.
*/
void *
__wt_lsm_checkpoint_worker(void *arg)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_LSM_WORKER_COOKIE cookie;
WT_SESSION_IMPL *session;
WT_TXN_ISOLATION saved_isolation;
u_int i, j;
int locked;
lsm_tree = arg;
session = lsm_tree->ckpt_session;
WT_CLEAR(cookie);
while (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING)) {
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret =
__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(ret);
}
WT_ERR(__lsm_copy_chunks(session, lsm_tree, &cookie, 0));
/* Write checkpoints in all completed files. */
for (i = 0, j = 0; i < cookie.nchunks - 1; i++) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_WORKING))
goto err;
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
break;
chunk = cookie.chunk_array[i];
/* Stop if a running transaction needs the chunk. */
__wt_txn_update_oldest(session);
if (!__wt_txn_visible_all(session, chunk->txnid_max))
break;
/*
* If the chunk is already checkpointed, make sure it
* is also evicted. Either way, there is no point
* trying to checkpoint it again.
*/
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_ONDISK)) {
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_EVICTED))
continue;
if ((ret = __lsm_discard_handle(
session, chunk->uri, NULL)) == 0)
F_SET_ATOMIC(
chunk, WT_LSM_CHUNK_EVICTED);
else if (ret == EBUSY)
ret = 0;
else
WT_ERR_MSG(session, ret,
"discard handle");
continue;
}
WT_VERBOSE_ERR(session, lsm,
"LSM worker flushing %u", i);
/*
* Flush the file before checkpointing: this is the
* expensive part in terms of I/O: do it without
* holding the schema lock.
*
* Use the special eviction isolation level to avoid
* interfering with an application checkpoint: we have
* already checked that all of the updates in this
* chunk are globally visible.
*
* !!! We can wait here for checkpoints and fsyncs to
* complete, which can be a long time.
*
* Don't keep waiting for the lock if application
* threads are waiting for a switch. Don't skip
* flushing the leaves either: that just means we'll
* hold the schema lock for (much) longer, which blocks
* the world.
*/
WT_ERR(__wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0));
for (locked = 0;
!locked && ret == 0 &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH);) {
if ((ret = __wt_spin_trylock(session,
&S2C(session)->checkpoint_lock)) == 0)
locked = 1;
else if (ret == EBUSY) {
__wt_yield();
ret = 0;
}
}
if (locked) {
saved_isolation = session->txn.isolation;
session->txn.isolation = TXN_ISO_EVICTION;
ret = __wt_bt_cache_op(
session, NULL, WT_SYNC_WRITE_LEAVES);
session->txn.isolation = saved_isolation;
__wt_spin_unlock(
session, &S2C(session)->checkpoint_lock);
}
WT_TRET(__wt_session_release_btree(session));
WT_ERR(ret);
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
break;
WT_VERBOSE_ERR(session, lsm,
"LSM worker checkpointing %u", i);
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_worker(session, chunk->uri,
__wt_checkpoint, NULL, NULL, 0));
if (ret != 0) {
__wt_err(session, ret, "LSM checkpoint");
break;
}
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
/*
* Clear the "cache resident" flag so the primary can
* be evicted and eventually closed. Only do this once
* the checkpoint has succeeded: otherwise, accessing
* the leaf page during the checkpoint can trigger
* forced eviction.
*/
WT_ERR(__wt_session_get_btree(
session, chunk->uri, NULL, NULL, 0));
__wt_btree_evictable(session, 1);
WT_ERR(__wt_session_release_btree(session));
++j;
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
F_SET_ATOMIC(chunk, WT_LSM_CHUNK_ONDISK);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
/* Update the throttle time. */
__wt_lsm_tree_throttle(session, lsm_tree);
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
/* Make sure we aren't pinning a transaction ID. */
__wt_txn_release_snapshot(session);
if (ret != 0) {
__wt_err(session, ret,
"LSM checkpoint metadata write");
break;
}
WT_VERBOSE_ERR(session, lsm,
"LSM worker checkpointed %u", i);
}
__lsm_unpin_chunks(session, &cookie);
if (j == 0 && F_ISSET(lsm_tree, WT_LSM_TREE_WORKING) &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH))
WT_ERR_TIMEDOUT_OK(__wt_cond_wait(
session, lsm_tree->work_cond, 100000));
}
err: __lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
/*
* The thread will only exit with failure if we run out of memory or
* there is some other system driven failure. We can't keep going
* after such a failure - ensure WiredTiger shuts down.
*/
if (ret != 0 && ret != WT_NOTFOUND)
WT_PANIC_ERR(session, ret,
"Shutting down LSM checkpoint utility thread");
return (NULL);
}
/*
* __wt_lsm_merge_worker --
* The merge worker thread for an LSM tree, responsible for merging
* on-disk trees.
*/
void *
__wt_lsm_merge_worker(void *vargs)
{
WT_DECL_RET;
WT_LSM_WORKER_ARGS *args;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
u_int aggressive, chunk_wait, id, old_aggressive, stallms;
int progress;
args = vargs;
lsm_tree = args->lsm_tree;
id = args->id;
session = lsm_tree->worker_sessions[id];
__wt_free(session, args);
aggressive = stallms = 0;
while (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING)) {
/*
* Help out with switching chunks in case the checkpoint worker
* is busy.
*/
if (F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
WT_WITH_SCHEMA_LOCK(session, ret =
__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(ret);
}
progress = 0;
/* Clear any state from previous worker thread iterations. */
session->dhandle = NULL;
/* Try to create a Bloom filter. */
if (__lsm_bloom_work(session, lsm_tree) == 0)
progress = 1;
/* If we didn't create a Bloom filter, try to merge. */
if (progress == 0 &&
__wt_lsm_merge(session, lsm_tree, id, aggressive) == 0)
progress = 1;
/* Clear any state from previous worker thread iterations. */
WT_CLEAR_BTREE_IN_SESSION(session);
/*
* Only have one thread freeing old chunks, and only if there
* are chunks to free.
*/
if (id == 0 && lsm_tree->nold_chunks > 0 &&
__lsm_free_chunks(session, lsm_tree) == 0)
progress = 1;
if (progress)
stallms = 0;
else if (F_ISSET(lsm_tree, WT_LSM_TREE_WORKING) &&
!F_ISSET(lsm_tree, WT_LSM_TREE_NEED_SWITCH)) {
/* Poll 10 times per second. */
WT_ERR_TIMEDOUT_OK(__wt_cond_wait(
session, lsm_tree->work_cond, 100000));
stallms += 100;
/*
* Get aggressive if more than enough chunks for a
* merge should have been created while we waited.
* Use 10 seconds as a default if we don't have an
* estimate.
*/
chunk_wait = stallms / (lsm_tree->chunk_fill_ms == 0 ?
10000 : lsm_tree->chunk_fill_ms);
old_aggressive = aggressive;
aggressive = chunk_wait / lsm_tree->merge_min;
if (aggressive > old_aggressive)
WT_VERBOSE_ERR(session, lsm,
"LSM merge got aggressive (%u), "
"%u / %" PRIu64,
aggressive, stallms,
lsm_tree->chunk_fill_ms);
}
}
if (0) {
err: __wt_err(session, ret, "LSM merge worker failed");
}
return (NULL);
}
/*
* __session_truncate --
* WT_SESSION->truncate method.
*/
static int
__session_truncate(WT_SESSION *wt_session,
const char *uri, WT_CURSOR *start, WT_CURSOR *stop, const char *config)
{
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_CURSOR *cursor;
int cmp;
session = (WT_SESSION_IMPL *)wt_session;
SESSION_TXN_API_CALL(session, truncate, config, cfg);
/*
* If the URI is specified, we don't need a start/stop, if start/stop
* is specified, we don't need a URI.
*
* If no URI is specified, and both cursors are specified, start/stop
* must reference the same object.
*
* Any specified cursor must have been initialized.
*/
if ((uri == NULL && start == NULL && stop == NULL) ||
(uri != NULL && (start != NULL || stop != NULL)))
WT_ERR_MSG(session, EINVAL,
"the truncate method should be passed either a URI or "
"start/stop cursors, but not both");
if (uri != NULL) {
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_schema_truncate(session, uri, cfg));
goto done;
}
/* Truncate is only supported for file and table objects. */
cursor = start == NULL ? stop : start;
if (!WT_PREFIX_MATCH(cursor->uri, "file:") &&
!WT_PREFIX_MATCH(cursor->uri, "table:"))
WT_ERR(__wt_bad_object_type(session, cursor->uri));
/*
* If both cursors set, check they're correctly ordered with respect to
* each other. We have to test this before any search, the search can
* change the initial cursor position.
*
* Rather happily, the compare routine will also confirm the cursors
* reference the same object and the keys are set.
*/
if (start != NULL && stop != NULL) {
WT_ERR(start->compare(start, stop, &cmp));
if (cmp > 0)
WT_ERR_MSG(session, EINVAL,
"the start cursor position is after the stop "
"cursor position");
}
/*
* Truncate does not require keys actually exist so that applications
* can discard parts of the object's name space without knowing exactly
* what records currently appear in the object. For this reason, do a
* search-near, rather than a search. Additionally, we have to correct
* after calling search-near, to position the start/stop cursors on the
* next record greater than/less than the original key. If the cursors
* hit the beginning/end of the object, or the start/stop keys cross,
* we're done, the range must be empty.
*/
if (start != NULL) {
WT_ERR(start->search_near(start, &cmp));
if (cmp < 0 && (ret = start->next(start)) != 0) {
WT_ERR_NOTFOUND_OK(ret);
goto done;
}
}
if (stop != NULL) {
WT_ERR(stop->search_near(stop, &cmp));
if (cmp > 0 && (ret = stop->prev(stop)) != 0) {
WT_ERR_NOTFOUND_OK(ret);
goto done;
}
if (start != NULL) {
WT_ERR(start->compare(start, stop, &cmp));
if (cmp > 0)
goto done;
}
}
if (WT_PREFIX_MATCH(cursor->uri, "file:"))
WT_ERR(__wt_curfile_truncate(session, start, stop));
else
WT_ERR(__wt_curtable_truncate(session, start, stop));
done:
err: TXN_API_END_NOTFOUND_MAP(session, ret);
}
/*
* __clsm_enter --
* Start an operation on an LSM cursor, update if the tree has changed.
*/
static inline int
__clsm_enter(WT_CURSOR_LSM *clsm, int reset, int update)
{
WT_CURSOR *c;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_SESSION_IMPL *session;
uint64_t *txnid_maxp;
uint64_t id, myid, snap_min;
session = (WT_SESSION_IMPL *)clsm->iface.session;
/* Merge cursors never update. */
if (F_ISSET(clsm, WT_CLSM_MERGE))
return (0);
if (reset) {
c = &clsm->iface;
/* Copy out data before resetting chunk cursors. */
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));
if (F_ISSET(c, WT_CURSTD_VALUE_INT) &&
!WT_DATA_IN_ITEM(&c->value))
WT_RET(__wt_buf_set(
session, &c->value, c->value.data, c->value.size));
WT_RET(__clsm_reset_cursors(clsm, NULL));
}
for (;;) {
/*
* If the cursor looks up-to-date, check if the cache is full.
* In case this call blocks, the check will be repeated before
* proceeding.
*/
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
WT_RET(__wt_cache_full_check(session));
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
/* Update the maximum transaction ID in the primary chunk. */
if (update && (chunk = clsm->primary_chunk) != NULL) {
WT_RET(__wt_txn_autocommit_check(session));
for (id = chunk->txnid_max, myid = session->txn.id;
!TXNID_LE(myid, id);
id = chunk->txnid_max) {
WT_ASSERT(session, myid != WT_TXN_NONE);
(void)WT_ATOMIC_CAS(
chunk->txnid_max, id, myid);
}
}
/*
* Figure out how many updates are required for snapshot
* isolation.
*
* This is not a normal visibility check on the maximum
* transaction ID in each chunk: any transaction ID that
* overlaps with our snapshot is a potential conflict.
*/
clsm->nupdates = 1;
if (session->txn.isolation == TXN_ISO_SNAPSHOT &&
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
snap_min = session->txn.snap_min;
for (txnid_maxp = &clsm->txnid_max[clsm->nchunks - 2];
clsm->nupdates < clsm->nchunks;
clsm->nupdates++, txnid_maxp--)
if (TXNID_LT(*txnid_maxp, snap_min))
break;
}
/*
* Stop when we are up-to-date, as long as this is:
* - a snapshot isolation update and the cursor is set up for
* that;
* - an update operation with a primary chunk, or
* - a read operation and the cursor is open for reading.
*/
if ((!update ||
session->txn.isolation != TXN_ISO_SNAPSHOT ||
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) &&
((update && clsm->primary_chunk != NULL) ||
(!update && F_ISSET(clsm, WT_CLSM_OPEN_READ))))
break;
open: WT_WITH_SCHEMA_LOCK(session,
ret = __clsm_open_cursors(clsm, update, 0, 0));
WT_RET(ret);
}
if (!F_ISSET(clsm, WT_CLSM_ACTIVE)) {
WT_RET(__cursor_enter(session));
F_SET(clsm, WT_CLSM_ACTIVE);
}
return (0);
}