/*
* __wt_lsm_work_bloom --
* Try to create a Bloom filter for the newest on-disk chunk that doesn't
* have one.
*/
int
__wt_lsm_work_bloom(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i, merge;
WT_CLEAR(cookie);
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, 0));
/* Create bloom filters in all checkpointed chunks. */
merge = 0;
for (i = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
/*
* Skip if a thread is still active in the chunk or it
* isn't suitable.
*/
if (!F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) ||
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM | WT_LSM_CHUNK_MERGING) ||
chunk->generation > 0 ||
chunk->count == 0)
continue;
/*
* See if we win the race to switch on the "busy" flag and
* recheck that the chunk still needs a Bloom filter.
*/
if (WT_ATOMIC_CAS4(chunk->bloom_busy, 0, 1)) {
if (!F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
ret = __lsm_bloom_create(
session, lsm_tree, chunk, (u_int)i);
/*
* Record if we were successful so that we can
* later push a merge work unit.
*/
if (ret == 0)
merge = 1;
}
chunk->bloom_busy = 0;
break;
}
}
/*
* If we created any bloom filters, we push a merge work unit now.
*/
if (merge)
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_MERGE, 0, lsm_tree));
err:
__lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
return (ret);
}
/*
* __lsm_bloom_work --
* Try to create a Bloom filter for the newest on-disk chunk.
*/
static int
__lsm_bloom_work(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i;
WT_CLEAR(cookie);
/* If no work is done, tell our caller by returning WT_NOTFOUND. */
ret = WT_NOTFOUND;
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, 0));
/* Create bloom filters in all checkpointed chunks. */
for (i = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
/*
* Skip if a thread is still active in the chunk or it
* isn't suitable.
*/
if (!F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_ONDISK) ||
F_ISSET_ATOMIC(chunk,
WT_LSM_CHUNK_BLOOM | WT_LSM_CHUNK_MERGING) ||
chunk->generation > 0 ||
chunk->count == 0)
continue;
/* See if we win the race to switch on the "busy" flag. */
if (WT_ATOMIC_CAS(chunk->bloom_busy, 0, 1)) {
ret = __lsm_bloom_create(
session, lsm_tree, chunk, (u_int)i);
chunk->bloom_busy = 0;
break;
}
}
__lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
return (ret);
}
/*
* __wt_lsm_free_chunks --
* Try to drop chunks from the tree that are no longer required.
*/
int
__wt_lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i, skipped;
int drop_ret;
bool flush_metadata;
flush_metadata = false;
if (lsm_tree->nold_chunks == 0)
return (0);
/*
* Make sure only a single thread is freeing the old chunk array
* at any time.
*/
if (!__wt_atomic_cas32(&lsm_tree->freeing_old_chunks, 0, 1))
return (0);
/*
* Take a copy of the current state of the LSM tree and look for chunks
* to drop. We do it this way to avoid holding the LSM tree lock while
* doing I/O or waiting on the schema lock.
*
* This is safe because only one thread will be in this function at a
* time. Merges may complete concurrently, and the old_chunks array
* may be extended, but we shuffle down the pointers each time we free
* one to keep the non-NULL slots at the beginning of the array.
*/
WT_CLEAR(cookie);
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, true));
for (i = skipped = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
WT_ASSERT(session, chunk != NULL);
/* Skip the chunk if another worker is using it. */
if (chunk->refcnt > 1) {
++skipped;
continue;
}
/*
* Don't remove files if a hot backup is in progress.
*
* The schema lock protects the set of live files, this check
* prevents us from removing a file that hot backup already
* knows about.
*/
if (S2C(session)->hot_backup)
break;
/*
* Drop any bloom filters and chunks we can. Don't try to drop
* a chunk if the bloom filter drop fails.
* An EBUSY return indicates that a cursor is still open in
* the tree - move to the next chunk in that case.
* An ENOENT return indicates that the LSM tree metadata was
* out of sync with the on disk state. Update the
* metadata to match in that case.
*/
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) {
drop_ret = __lsm_drop_file(session, chunk->bloom_uri);
if (drop_ret == EBUSY) {
++skipped;
continue;
} else if (drop_ret != ENOENT)
WT_ERR(drop_ret);
flush_metadata = true;
F_CLR(chunk, WT_LSM_CHUNK_BLOOM);
}
if (chunk->uri != NULL) {
drop_ret = __lsm_drop_file(session, chunk->uri);
if (drop_ret == EBUSY) {
++skipped;
continue;
} else if (drop_ret != ENOENT)
WT_ERR(drop_ret);
flush_metadata = true;
}
/* Lock the tree to clear out the old chunk information. */
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
/*
* The chunk we are looking at should be the first one in the
* tree that we haven't already skipped over.
*/
WT_ASSERT(session, lsm_tree->old_chunks[skipped] == chunk);
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, lsm_tree->old_chunks[skipped]);
/* Shuffle down to keep all occupied slots at the beginning. */
if (--lsm_tree->nold_chunks > skipped) {
memmove(lsm_tree->old_chunks + skipped,
lsm_tree->old_chunks + skipped + 1,
(lsm_tree->nold_chunks - skipped) *
sizeof(WT_LSM_CHUNK *));
lsm_tree->old_chunks[lsm_tree->nold_chunks] = NULL;
}
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
/*
* Clear the chunk in the cookie so we don't attempt to
* decrement the reference count.
*/
cookie.chunk_array[i] = NULL;
}
err: /* Flush the metadata unless the system is in panic */
if (flush_metadata && ret != WT_PANIC) {
WT_TRET(__wt_lsm_tree_writelock(session, lsm_tree));
WT_TRET(__wt_lsm_meta_write(session, lsm_tree));
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
}
__lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
lsm_tree->freeing_old_chunks = 0;
/* Returning non-zero means there is no work to do. */
if (!flush_metadata)
WT_TRET(WT_NOTFOUND);
return (ret);
}
/*
* __lsm_free_chunks --
* Try to drop chunks from the tree that are no longer required.
*/
static int
__lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_WORKER_COOKIE cookie;
u_int i, skipped;
int progress;
/*
* Take a copy of the current state of the LSM tree and look for chunks
* to drop. We do it this way to avoid holding the LSM tree lock while
* doing I/O or waiting on the schema lock.
*
* This is safe because only one thread will be in this function at a
* time (the first merge thread). Merges may complete concurrently,
* and the old_chunks array may be extended, but we shuffle down the
* pointers each time we free one to keep the non-NULL slots at the
* beginning of the array.
*/
WT_CLEAR(cookie);
WT_RET(__lsm_copy_chunks(session, lsm_tree, &cookie, 1));
for (i = skipped = 0, progress = 0; i < cookie.nchunks; i++) {
chunk = cookie.chunk_array[i];
WT_ASSERT(session, chunk != NULL);
/* Skip the chunk if another worker is using it. */
if (chunk->refcnt > 1) {
++skipped;
continue;
}
if (F_ISSET_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM)) {
/*
* An EBUSY return is acceptable - a cursor may still
* be positioned on this old chunk.
*/
if ((ret = __lsm_drop_file(
session, chunk->bloom_uri)) == EBUSY) {
WT_VERBOSE_ERR(session, lsm,
"LSM worker bloom drop busy: %s.",
chunk->bloom_uri);
++skipped;
continue;
} else
WT_ERR(ret);
F_CLR_ATOMIC(chunk, WT_LSM_CHUNK_BLOOM);
}
if (chunk->uri != NULL) {
/*
* An EBUSY return is acceptable - a cursor may still
* be positioned on this old chunk.
*/
if ((ret = __lsm_drop_file(
session, chunk->uri)) == EBUSY) {
WT_VERBOSE_ERR(session, lsm,
"LSM worker drop busy: %s.",
chunk->uri);
++skipped;
continue;
} else
WT_ERR(ret);
}
progress = 1;
/* Lock the tree to clear out the old chunk information. */
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 1));
/*
* The chunk we are looking at should be the first one in the
* tree that we haven't already skipped over.
*/
WT_ASSERT(session, lsm_tree->old_chunks[skipped] == chunk);
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, lsm_tree->old_chunks[skipped]);
/* Shuffle down to keep all occupied slots at the beginning. */
if (--lsm_tree->nold_chunks > skipped) {
memmove(lsm_tree->old_chunks + skipped,
lsm_tree->old_chunks + skipped + 1,
(lsm_tree->nold_chunks - skipped) *
sizeof(WT_LSM_CHUNK *));
lsm_tree->old_chunks[lsm_tree->nold_chunks] = NULL;
}
/*
* Clear the chunk in the cookie so we don't attempt to
* decrement the reference count.
*/
cookie.chunk_array[i] = NULL;
/*
* Update the metadata. We used to try to optimize by only
* updating the metadata once at the end, but the error
* handling is not straightforward.
*/
WT_TRET(__wt_lsm_meta_write(session, lsm_tree));
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
}
err: __lsm_unpin_chunks(session, &cookie);
__wt_free(session, cookie.chunk_array);
/* Returning non-zero means there is no work to do. */
if (!progress)
WT_TRET(WT_NOTFOUND);
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);
}
