/*
* __lsm_bloom_create --
* Create a bloom filter for a chunk of the LSM tree that has been
* checkpointed but not yet been merged.
*/
static int
__lsm_bloom_create(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk, u_int chunk_off)
{
WT_BLOOM *bloom;
WT_CURSOR *src;
WT_DECL_RET;
WT_ITEM key;
uint64_t insert_count;
WT_RET(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
bloom = NULL;
/*
* This is merge-like activity, and we don't want compacts to give up
* because we are creating a bunch of bloom filters before merging.
*/
++lsm_tree->merge_progressing;
WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config, chunk->count,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
/* Open a special merge cursor just on this chunk. */
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, chunk_off, chunk->id, 1));
/*
* Setup so that we don't hold pages we read into cache, and so
* that we don't get stuck if the cache is full. If we allow
* ourselves to get stuck creating bloom filters, the entire tree
* can stall since there may be no worker threads available to flush.
*/
F_SET(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
WT_ERR(src->get_key(src, &key));
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_TRET(src->close(src));
WT_TRET(__wt_bloom_finalize(bloom));
WT_ERR(ret);
F_CLR(session, WT_SESSION_NO_CACHE);
/* Load the new Bloom filter into cache. */
WT_CLEAR(key);
WT_ERR_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"LSM worker created bloom filter %s. "
"Expected %" PRIu64 " items, got %" PRIu64,
chunk->bloom_uri, chunk->count, insert_count));
/* Ensure the bloom filter is in the metadata. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM bloom worker metadata write");
err: if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, int update, u_int start_chunk, uint32_t start_id)
{
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
int locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
lsm_tree = clsm->lsm_tree;
chunk = NULL;
ckpt_cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/* Copy the key, so we don't lose the cursor position. */
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));
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
if (update) {
if (txn->isolation == TXN_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
} else
F_SET(clsm, WT_CLSM_OPEN_READ);
WT_RET(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
/*
* If there is no in-memory chunk in the tree for an update operation,
* create one.
*
* !!!
* It is exceeding unlikely that we get here at all, but if there is a
* transaction in progress and it rolls back, it would leave the
* metadata inconsistent.
*/
if (update && (lsm_tree->nchunks == 0 ||
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) == NULL ||
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))) {
/* Release our lock because switch will get a write lock. */
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
}
F_SET(session, WT_SESSION_NO_CACHE_CHECK);
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
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++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->txnid_max));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->txnid_max[ngood - 1] =
chunk->txnid_max;
if (__wt_txn_visible_all(
session, chunk->txnid_max))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (clsm->cursors != NULL && (ngood < clsm->nchunks ||
(!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0))) {
saved_gen = lsm_tree->dsk_gen;
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0)
WT_ERR(__clsm_close_cursors(
clsm, 0, nchunks - nupdates));
WT_ERR(__clsm_close_cursors(
clsm, ngood, clsm->nchunks));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->txnid_max[i] = chunk->txnid_max;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL && !F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
WT_WITH_BTREE(session, ((WT_CURSOR_BTREE *)(primary))->btree,
__wt_btree_evictable(session, 0));
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err: F_CLR(session, WT_SESSION_NO_CACHE_CHECK);
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
return (ret);
}
/*
* __lsm_bloom_create --
* Create a bloom filter for a chunk of the LSM tree that has been
* checkpointed but not yet been merged.
*/
static int
__lsm_bloom_create(WT_SESSION_IMPL *session,
WT_LSM_TREE *lsm_tree, WT_LSM_CHUNK *chunk, u_int chunk_off)
{
WT_BLOOM *bloom;
WT_CURSOR *src;
WT_DECL_RET;
WT_ITEM buf, key;
WT_SESSION *wt_session;
uint64_t insert_count;
int exist;
/*
* Normally, the Bloom URI is populated when the chunk struct is
* allocated. After an open, however, it may not have been.
* Deal with that here.
*/
if (chunk->bloom_uri == NULL) {
WT_CLEAR(buf);
WT_RET(__wt_lsm_tree_bloom_name(
session, lsm_tree, chunk->id, &buf));
chunk->bloom_uri = __wt_buf_steal(session, &buf, NULL);
}
/*
* Drop the bloom filter first - there may be some content hanging over
* from an aborted merge or checkpoint.
*/
wt_session = &session->iface;
WT_RET(__wt_exist(session, chunk->bloom_uri + strlen("file:"), &exist));
if (exist)
WT_RET(wt_session->drop(wt_session, chunk->bloom_uri, "force"));
bloom = NULL;
/*
* This is merge-like activity, and we don't want compacts to give up
* because we are creating a bunch of bloom filters before merging.
*/
++lsm_tree->merge_progressing;
WT_RET(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config, chunk->count,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count, &bloom));
/* Open a special merge cursor just on this chunk. */
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, chunk_off, chunk->id, 1));
F_SET(session, WT_SESSION_NO_CACHE);
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
WT_ERR(src->get_key(src, &key));
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_TRET(src->close(src));
WT_TRET(__wt_bloom_finalize(bloom));
WT_ERR(ret);
F_CLR(session, WT_SESSION_NO_CACHE);
/* Load the new Bloom filter into cache. */
WT_CLEAR(key);
WT_ERR_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
WT_VERBOSE_ERR(session, lsm,
"LSM worker created bloom filter %s. "
"Expected %" PRIu64 " items, got %" PRIu64,
chunk->bloom_uri, chunk->count, insert_count);
/* Ensure the bloom filter is in the metadata. */
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
F_SET_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM);
ret = __wt_lsm_meta_write(session, lsm_tree);
++lsm_tree->dsk_gen;
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
if (ret != 0)
WT_ERR_MSG(session, ret, "LSM bloom worker metadata write");
err: if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
F_CLR(session, WT_SESSION_NO_CACHE);
return (ret);
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, bool update, u_int start_chunk, uint32_t start_id)
{
WT_BTREE *btree;
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
u_int close_range_end, close_range_start;
bool locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
chunk = NULL;
locked = false;
lsm_tree = clsm->lsm_tree;
/*
* Ensure that any snapshot update has cursors on the right set of
* chunks to guarantee visibility is correct.
*/
if (update && txn->isolation == WT_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
/*
* Query operations need a full set of cursors. Overwrite cursors
* do queries in service of updates.
*/
if (!update || !F_ISSET(c, WT_CURSTD_OVERWRITE))
F_SET(clsm, WT_CLSM_OPEN_READ);
if (lsm_tree->nchunks == 0)
return (0);
ckpt_cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/*
* If the key is pointing to memory that is pinned by a chunk
* cursor, take a copy before closing cursors.
*/
if (F_ISSET(c, WT_CURSTD_KEY_INT))
WT_CURSOR_NEEDKEY(c);
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
WT_RET(__wt_lsm_tree_readlock(session, lsm_tree));
locked = true;
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
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++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->switch_txn));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->switch_txn[ngood - 1] = chunk->switch_txn;
if (__wt_txn_visible_all(
session, chunk->switch_txn))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (clsm->cursors != NULL && ngood < clsm->nchunks) {
close_range_start = ngood;
close_range_end = clsm->nchunks;
} else if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0 ) {
close_range_start = 0;
close_range_end = WT_MIN(nchunks, clsm->nchunks);
if (close_range_end > nupdates)
close_range_end -= nupdates;
else
close_range_end = 0;
WT_ASSERT(session, ngood >= close_range_end);
} else {
close_range_end = 0;
close_range_start = 0;
}
if (close_range_end > close_range_start) {
saved_gen = lsm_tree->dsk_gen;
locked = false;
WT_ERR(__wt_lsm_tree_readunlock(session, lsm_tree));
WT_ERR(__clsm_close_cursors(
clsm, close_range_start, close_range_end));
WT_ERR(__wt_lsm_tree_readlock(session, lsm_tree));
locked = true;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->switch_txn[i] = chunk->switch_txn;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND && F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
/*
* Setup all cursors other than the primary to only do conflict
* checks on insert operations. This allows us to execute
* inserts on non-primary chunks as a way of checking for
* write conflicts with concurrent updates.
*/
if (i != nchunks - 1)
(*cp)->insert = __wt_curfile_update_check;
if (!F_ISSET(clsm, WT_CLSM_MERGE) &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL &&
!F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
chunk->switch_txn == WT_TXN_NONE) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
/*
* Disable eviction for the in-memory chunk. Also clear the
* bulk load flag here, otherwise eviction will be enabled by
* the first update.
*/
btree = ((WT_CURSOR_BTREE *)(primary))->btree;
if (btree->bulk_load_ok) {
btree->bulk_load_ok = false;
WT_WITH_BTREE(session, btree,
__wt_btree_evictable(session, false));
}
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err:
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_readunlock(session, lsm_tree));
return (ret);
}
/*
* __wt_curjoin_join --
* Add a new join to a join cursor.
*/
int
__wt_curjoin_join(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_INDEX *idx, WT_CURSOR *ref_cursor, uint8_t flags, uint8_t range,
uint64_t count, uint32_t bloom_bit_count, uint32_t bloom_hash_count)
{
WT_CURSOR_JOIN_ENTRY *entry;
WT_DECL_RET;
WT_CURSOR_JOIN_ENDPOINT *end, *newend;
bool hasins, needbloom, range_eq;
u_int i, ins, nonbloom;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
char *main_uri;
size_t namesize, newsize;
entry = NULL;
hasins = needbloom = false;
ins = 0; /* -Wuninitialized */
main_uri = NULL;
nonbloom = 0; /* -Wuninitialized */
namesize = strlen(cjoin->table->name);
for (i = 0; i < cjoin->entries_next; i++) {
if (cjoin->entries[i].index == idx) {
entry = &cjoin->entries[i];
break;
}
if (!needbloom && i > 0 &&
!F_ISSET(&cjoin->entries[i], WT_CURJOIN_ENTRY_BLOOM)) {
needbloom = true;
nonbloom = i;
}
}
if (entry == NULL) {
WT_ERR(__wt_realloc_def(session, &cjoin->entries_allocated,
cjoin->entries_next + 1, &cjoin->entries));
if (LF_ISSET(WT_CURJOIN_ENTRY_BLOOM) && needbloom) {
/*
* Reorder the list so that after the first entry,
* the Bloom filtered entries come next, followed by
* the non-Bloom entries. Once the Bloom filters
* are built, determining membership via Bloom is
* faster than without Bloom, so we can answer
* membership questions more quickly, and with less
* I/O, with the Bloom entries first.
*/
entry = &cjoin->entries[nonbloom];
memmove(entry + 1, entry,
(cjoin->entries_next - nonbloom) *
sizeof(WT_CURSOR_JOIN_ENTRY));
memset(entry, 0, sizeof(WT_CURSOR_JOIN_ENTRY));
}
else
entry = &cjoin->entries[cjoin->entries_next];
entry->index = idx;
entry->flags = flags;
entry->count = count;
entry->bloom_bit_count = bloom_bit_count;
entry->bloom_hash_count = bloom_hash_count;
++cjoin->entries_next;
} else {
/* Merge the join into an existing entry for this index */
if (count != 0 && entry->count != 0 && entry->count != count)
WT_ERR_MSG(session, EINVAL,
"count=%" PRIu64 " does not match "
"previous count=%" PRIu64 " for this index",
count, entry->count);
if (LF_MASK(WT_CURJOIN_ENTRY_BLOOM) !=
F_MASK(entry, WT_CURJOIN_ENTRY_BLOOM))
WT_ERR_MSG(session, EINVAL,
"join has incompatible strategy "
"values for the same index");
/*
* Check against other comparisons (we call them endpoints)
* already set up for this index.
* We allow either:
* - one or more "eq" (with disjunction)
* - exactly one "eq" (with conjunction)
* - exactly one of "gt" or "ge" (conjunction or disjunction)
* - exactly one of "lt" or "le" (conjunction or disjunction)
* - one of "gt"/"ge" along with one of "lt"/"le"
* (currently restricted to conjunction).
*
* Some other combinations, although expressible either do
* not make sense (X == 3 AND X == 5) or are reducible (X <
* 7 AND X < 9). Other specific cases of (X < 7 OR X > 15)
* or (X == 4 OR X > 15) make sense but we don't handle yet.
*/
for (i = 0; i < entry->ends_next; i++) {
end = &entry->ends[i];
range_eq = (range == WT_CURJOIN_END_EQ);
if ((F_ISSET(end, WT_CURJOIN_END_GT) &&
((range & WT_CURJOIN_END_GT) != 0 || range_eq)) ||
(F_ISSET(end, WT_CURJOIN_END_LT) &&
((range & WT_CURJOIN_END_LT) != 0 || range_eq)) ||
(end->flags == WT_CURJOIN_END_EQ &&
(range & (WT_CURJOIN_END_LT | WT_CURJOIN_END_GT))
!= 0))
WT_ERR_MSG(session, EINVAL,
"join has overlapping ranges");
if (range == WT_CURJOIN_END_EQ &&
end->flags == WT_CURJOIN_END_EQ &&
!F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
WT_ERR_MSG(session, EINVAL,
"compare=eq can only be combined "
"using operation=or");
/*
* Sort "gt"/"ge" to the front, followed by any number
* of "eq", and finally "lt"/"le".
*/
if (!hasins &&
((range & WT_CURJOIN_END_GT) != 0 ||
(range == WT_CURJOIN_END_EQ &&
!F_ISSET(end, WT_CURJOIN_END_GT)))) {
ins = i;
hasins = true;
}
}
/* All checks completed, merge any new configuration now */
entry->count = count;
entry->bloom_bit_count =
WT_MAX(entry->bloom_bit_count, bloom_bit_count);
entry->bloom_hash_count =
WT_MAX(entry->bloom_hash_count, bloom_hash_count);
}
WT_ERR(__wt_realloc_def(session, &entry->ends_allocated,
entry->ends_next + 1, &entry->ends));
if (!hasins)
ins = entry->ends_next;
newend = &entry->ends[ins];
memmove(newend + 1, newend,
(entry->ends_next - ins) * sizeof(WT_CURSOR_JOIN_ENDPOINT));
memset(newend, 0, sizeof(WT_CURSOR_JOIN_ENDPOINT));
entry->ends_next++;
newend->cursor = ref_cursor;
F_SET(newend, range);
/* Open the main file with a projection of the indexed columns. */
if (entry->main == NULL && entry->index != NULL) {
namesize = strlen(cjoin->table->name);
newsize = namesize + entry->index->colconf.len + 1;
WT_ERR(__wt_calloc(session, 1, newsize, &main_uri));
snprintf(main_uri, newsize, "%s%.*s",
cjoin->table->name, (int)entry->index->colconf.len,
entry->index->colconf.str);
WT_ERR(__wt_open_cursor(session, main_uri,
(WT_CURSOR *)cjoin, raw_cfg, &entry->main));
}
err: if (main_uri != NULL)
__wt_free(session, main_uri);
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_INDEX *cindex;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_RET;
WT_DECL_ITEM(uribuf);
WT_ITEM curkey, curvalue, *k;
WT_TABLE *maintable;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
const char *mainkey_str, *p;
void *allocbuf;
size_t mainkey_len, size;
u_int i;
int cmp, skip;
c = NULL;
allocbuf = NULL;
skip = 0;
if (entry->index != NULL) {
/*
* Open a cursor having a projection of the keys of the
* index we're comparing against. Open it raw, we're
* going to compare it to the raw keys of the
* reference cursors.
*/
maintable = ((WT_CURSOR_TABLE *)entry->main)->table;
mainkey_str = maintable->colconf.str + 1;
for (p = mainkey_str, i = 0;
p != NULL && i < maintable->nkey_columns; i++)
p = strchr(p + 1, ',');
WT_ASSERT(session, p != 0);
mainkey_len = WT_PTRDIFF(p, mainkey_str);
size = strlen(entry->index->name) + mainkey_len + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s(%.*s)",
entry->index->name, (int)mainkey_len, mainkey_str));
} 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->name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->name));
}
WT_ERR(__wt_open_cursor(
session, uribuf->data, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax &&
F_ISSET(end, WT_CURJOIN_END_GE)) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (end->flags == WT_CURJOIN_END_GE)
skip = 1;
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
if (entry->index != NULL) {
cindex = (WT_CURSOR_INDEX *)c;
if (cindex->index->extractor == NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
k = &cindex->child->key;
WT_ERR(__wt_struct_repack(session,
cindex->child->key_format,
entry->main->value_format, k, &curkey,
&allocbuf));
} else
curkey = cindex->child->key;
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
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;
}
}
if (entry->index != NULL)
WT_ERR(c->get_value(c, &curvalue));
else
WT_ERR(c->get_key(c, &curvalue));
WT_ERR(__wt_bloom_insert(bloom, &curvalue));
entry->stats.actual_count++;
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);
__wt_free(session, allocbuf);
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;
u_int created_chunk, verb;
int create_bloom, locked, in_sync, tret;
const char *cfg[3];
const char *drop_cfg[] =
{ WT_CONFIG_BASE(session, session_drop), "force", NULL };
bloom = NULL;
chunk = NULL;
create_bloom = 0;
created_chunk = 0;
dest = src = NULL;
locked = 0;
start_id = 0;
in_sync = 0;
/* Fast path if it's obvious no merges could be done. */
if (lsm_tree->nchunks < lsm_tree->merge_min &&
lsm_tree->merge_aggressiveness < WT_LSM_AGGRESSIVE_THRESHOLD)
return (WT_NOTFOUND);
/*
* Use the lsm_tree lock to read the chunks (so no switches occur), but
* avoid holding it while the merge is in progress: that may take a
* long time.
*/
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
locked = 1;
WT_ERR(__lsm_merge_span(session,
lsm_tree, id, &start_chunk, &end_chunk, &record_count));
nchunks = (end_chunk + 1) - start_chunk;
WT_ASSERT(session, nchunks > 0);
start_id = lsm_tree->chunk[start_chunk]->id;
/* Find the merge generation. */
for (generation = 0, i = 0; i < nchunks; i++)
generation = WT_MAX(generation,
lsm_tree->chunk[start_chunk + i]->generation + 1);
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
locked = 0;
/* Allocate an ID for the merge. */
dest_id = WT_ATOMIC_ADD4(lsm_tree->last, 1);
/*
* We only want to do the chunk loop if we're running with verbose,
* so we wrap these statements in the conditional. Avoid the loop
* in the normal path.
*/
if (WT_VERBOSE_ISSET(session, WT_VERB_LSM)) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s chunks %u-%u into %u (%" PRIu64 " records)"
", generation %" PRIu32,
lsm_tree->name,
start_chunk, end_chunk, dest_id, record_count, generation));
for (verb = start_chunk; verb <= end_chunk; verb++)
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"%s: Chunk[%u] id %u",
lsm_tree->name, verb, lsm_tree->chunk[verb]->id));
}
WT_ERR(__wt_calloc_one(session, &chunk));
created_chunk = 1;
chunk->id = dest_id;
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_MERGED) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST) ||
start_chunk > 0) && record_count > 0)
create_bloom = 1;
/*
* Special setup for the merge cursor:
* first, reset to open the dependent cursors;
* then restrict the cursor to a specific number of chunks;
* then set MERGE so the cursor doesn't track updates to the tree.
*/
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, start_chunk, start_id, nchunks));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
WT_ERR(ret);
if (create_bloom) {
WT_ERR(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
WT_ERR(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config,
record_count, lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count, &bloom));
}
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
cfg[1] = "bulk,raw,skip_sort_check";
cfg[2] = NULL;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
#define LSM_MERGE_CHECK_INTERVAL 1000
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
if (insert_count % LSM_MERGE_CHECK_INTERVAL == 0) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE))
WT_ERR(EINTR);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
}
WT_ERR(src->get_key(src, &key));
dest->set_key(dest, &key);
WT_ERR(src->get_value(src, &value));
dest->set_value(dest, &value);
WT_ERR(dest->insert(dest));
if (create_bloom)
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, insert_count % LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Bloom size for %" PRIu64 " has %" PRIu64 " items inserted.",
record_count, insert_count));
/*
* Closing and syncing the files can take a while. Set the
* merge_syncing field so that compact knows it is still in
* progress.
*/
(void)WT_ATOMIC_ADD4(lsm_tree->merge_syncing, 1);
in_sync = 1;
/*
* We've successfully created the new chunk. Now install it. We need
* to ensure that the NO_CACHE flag is cleared and the bloom filter
* is closed (even if a step fails), so track errors but don't return
* until we've cleaned up.
*/
WT_TRET(src->close(src));
WT_TRET(dest->close(dest));
src = dest = NULL;
F_CLR(session, WT_SESSION_NO_CACHE);
/*
* We're doing advisory reads to fault the new trees into cache.
* Don't block if the cache is full: our next unit of work may be to
* discard some trees to free space.
*/
F_SET(session, WT_SESSION_NO_CACHE_CHECK);
if (create_bloom) {
if (ret == 0)
WT_TRET(__wt_bloom_finalize(bloom));
/*
* Read in a key to make sure the Bloom filters btree handle is
* open before it becomes visible to application threads.
* Otherwise application threads will stall while it is opened
* and internal pages are read into cache.
*/
if (ret == 0) {
WT_CLEAR(key);
WT_TRET_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
}
WT_TRET(__wt_bloom_close(bloom));
bloom = NULL;
}
WT_ERR(ret);
/*
* Open a handle on the new chunk before application threads attempt
* to access it, opening it pre-loads internal pages into the file
* system cache.
*/
cfg[1] = "checkpoint=" WT_CHECKPOINT;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
WT_TRET(dest->close(dest));
dest = NULL;
++lsm_tree->merge_progressing;
(void)WT_ATOMIC_SUB4(lsm_tree->merge_syncing, 1);
in_sync = 0;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = 1;
/*
* Check whether we raced with another merge, and adjust the chunk
* array offset as necessary.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id)
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++)
if (lsm_tree->chunk[start_chunk]->id == start_id)
break;
/*
* It is safe to error out here - since the update can only fail
* prior to making updates to the tree.
*/
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, start_chunk, nchunks, chunk));
if (create_bloom)
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
chunk->count = insert_count;
chunk->generation = generation;
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
/*
* We have no current way of continuing if the metadata update fails,
* so we will panic in that case. Put some effort into cleaning up
* after ourselves here - so things have a chance of shutting down.
*
* Any errors that happened after the tree was locked are
* fatal - we can't guarantee the state of the tree.
*/
if ((ret = __wt_lsm_meta_write(session, lsm_tree)) != 0)
WT_PANIC_ERR(session, ret, "Failed finalizing LSM merge");
lsm_tree->dsk_gen++;
/* Update the throttling while holding the tree lock. */
__wt_lsm_tree_throttle(session, lsm_tree, 1);
/* Schedule a pass to discard old chunks */
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_DROP, 0, lsm_tree));
err: if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (in_sync)
(void)WT_ATOMIC_SUB4(lsm_tree->merge_syncing, 1);
if (src != NULL)
WT_TRET(src->close(src));
if (dest != NULL)
WT_TRET(dest->close(dest));
if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
if (ret != 0 && created_chunk) {
/* Drop the newly-created files on error. */
if (chunk->uri != NULL) {
WT_WITH_SCHEMA_LOCK(session, tret =
__wt_schema_drop(session, chunk->uri, drop_cfg));
WT_TRET(tret);
}
if (create_bloom && chunk->bloom_uri != NULL) {
WT_WITH_SCHEMA_LOCK(session,
tret = __wt_schema_drop(
session, chunk->bloom_uri, drop_cfg));
WT_TRET(tret);
}
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, chunk);
if (ret == EINTR)
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge aborted due to close"));
else
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge failed with %s", wiredtiger_strerror(ret)));
}
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_CACHE_CHECK);
return (ret);
}
