/*
* __cursor_key_order_check_row --
* Check key ordering for row-store cursor movements.
*/
static int
__cursor_key_order_check_row(
WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, bool next)
{
WT_BTREE *btree;
WT_ITEM *key;
WT_DECL_RET;
WT_DECL_ITEM(a);
WT_DECL_ITEM(b);
int cmp;
btree = S2BT(session);
key = &cbt->iface.key;
cmp = 0; /* -Werror=maybe-uninitialized */
if (cbt->lastkey->size != 0)
WT_RET(__wt_compare(
session, btree->collator, cbt->lastkey, key, &cmp));
if (cbt->lastkey->size == 0 || (next && cmp < 0) || (!next && cmp > 0))
return (__wt_buf_set(session,
cbt->lastkey, cbt->iface.key.data, cbt->iface.key.size));
WT_ERR(__wt_scr_alloc(session, 512, &a));
WT_ERR(__wt_scr_alloc(session, 512, &b));
WT_PANIC_ERR(session, EINVAL,
"WT_CURSOR.%s out-of-order returns: returned key %s then key %s",
next ? "next" : "prev",
__wt_buf_set_printable(
session, cbt->lastkey->data, cbt->lastkey->size, a),
__wt_buf_set_printable(session, key->data, key->size, b));
err: __wt_scr_free(session, &a);
__wt_scr_free(session, &b);
return (ret);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_ITEM curkey, curvalue;
size_t size;
u_int skip;
int cmp;
const char *uri;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
c = NULL;
skip = 0;
if (entry->index != NULL)
/*
* Open the raw index. We're avoiding any references
* to the main table, they may be expensive.
*/
uri = entry->index->source;
else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->iface.name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->iface.name));
uri = uribuf->data;
}
WT_ERR(__wt_open_cursor(session, uri, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax) {
if (F_ISSET(end, WT_CURJOIN_END_GT) ||
WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_EQ) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_GE)
skip = 1;
} else if (F_ISSET(end, WT_CURJOIN_END_LT)) {
if ((ret = c->next(c)) == WT_NOTFOUND)
goto done;
WT_ERR(ret);
} else
WT_PANIC_ERR(session, EINVAL,
"fatal error in join cursor position state");
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
entry->stats.iterated++;
if (entry->index != NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
WT_ERR(__wt_struct_repack(session,
c->key_format,
(entry->repack_format != NULL ?
entry->repack_format : entry->index->idxkey_format),
&c->key, &curkey));
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION)) {
/* if condition satisfied, insert immediately */
switch (WT_CURJOIN_END_RANGE(end)) {
case WT_CURJOIN_END_EQ:
if (cmp == 0)
goto insert;
break;
case WT_CURJOIN_END_GT:
if (cmp > 0) {
/* skip this check next time */
skip = entry->ends_next;
goto insert;
}
break;
case WT_CURJOIN_END_GE:
if (cmp >= 0)
goto insert;
break;
case WT_CURJOIN_END_LT:
if (cmp < 0)
goto insert;
break;
case WT_CURJOIN_END_LE:
if (cmp <= 0)
goto insert;
break;
}
} else if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
/*
* Either it's a disjunction that hasn't satisfied any
* condition, or it's a conjunction that has satisfied all
* conditions.
*/
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
goto advance;
insert:
if (entry->index != NULL) {
curvalue.data =
(unsigned char *)curkey.data + curkey.size;
WT_ASSERT(session, c->key.size > curkey.size);
curvalue.size = c->key.size - curkey.size;
}
else
WT_ERR(c->get_key(c, &curvalue));
__wt_bloom_insert(bloom, &curvalue);
entry->stats.bloom_insert++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
return (ret);
}
/*
* __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);
}
/*
* __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);
}
