/*
* __truncate_table --
* WT_SESSION::truncate for a table.
*/
static int
__truncate_table(WT_SESSION_IMPL *session, const char *name)
{
WT_BTREE *btree;
WT_DECL_ITEM(namebuf);
WT_DECL_RET;
WT_TABLE *table;
int i;
WT_RET(__wt_schema_get_table(session, name, strlen(name), 0, &table));
WT_RET(__wt_scr_alloc(session, 0, &namebuf));
/* Truncate the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->cgroups[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
/* Truncate the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->indices[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
table->idx_complete = 0;
/* Reopen the column groups. */
ret = __wt_schema_open_colgroups(session, table);
err: __wt_scr_free(&namebuf);
return (ret);
}
/*
* __wt_metadata_open --
* Opens the metadata file, sets session->meta_dhandle.
*/
int
__wt_metadata_open(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
if (session->meta_dhandle != NULL)
return (0);
WT_RET(__wt_session_get_btree(session, WT_METAFILE_URI, NULL, NULL, 0));
session->meta_dhandle = session->dhandle;
WT_ASSERT(session, session->meta_dhandle != NULL);
/*
* Set special flags for the metadata file: eviction (the metadata file
* is in-memory and never evicted), logging (the metadata file is always
* logged if possible).
*
* Test flags before setting them so updates can't race in subsequent
* opens (the first update is safe because it's single-threaded from
* wiredtiger_open).
*/
btree = S2BT(session);
if (!F_ISSET(btree, WT_BTREE_IN_MEMORY))
F_SET(btree, WT_BTREE_IN_MEMORY);
if (!F_ISSET(btree, WT_BTREE_NO_EVICTION))
F_SET(btree, WT_BTREE_NO_EVICTION);
if (F_ISSET(btree, WT_BTREE_NO_LOGGING))
F_CLR(btree, WT_BTREE_NO_LOGGING);
/* The metadata handle doesn't need to stay locked -- release it. */
return (__wt_session_release_btree(session));
}
/*
* __truncate_file --
* WT_SESSION::truncate for a file.
*/
static int
__truncate_file(WT_SESSION_IMPL *session, const char *name)
{
WT_DECL_RET;
const char *filename;
uint32_t allocsize;
filename = name;
if (!WT_PREFIX_SKIP(filename, "file:"))
return (EINVAL);
/* Open and lock the file. */
WT_RET(__wt_session_get_btree(
session, name, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
/* Get the allocation size. */
allocsize = S2BT(session)->allocsize;
WT_RET(__wt_session_release_btree(session));
/* Close any btree handles in the file. */
WT_WITH_DHANDLE_LOCK(session,
ret = __wt_conn_dhandle_close_all(session, name, 0));
WT_RET(ret);
/* Delete the root address and truncate the file. */
WT_RET(__wt_meta_checkpoint_clear(session, name));
WT_RET(__wt_block_manager_truncate(session, filename, allocsize));
return (0);
}
/*
* __conn_btree_apply_internal --
* Apply a function to the open btree handles.
*/
static int
__conn_btree_apply_internal(WT_SESSION_IMPL *session, WT_DATA_HANDLE *dhandle,
int (*func)(WT_SESSION_IMPL *, const char *[]), const char *cfg[])
{
WT_DECL_RET;
/*
* We need to pull the handle into the session handle
* cache and make sure it's referenced to stop other
* internal code dropping the handle (e.g in LSM when
* cleaning up obsolete chunks). Holding the metadata
* lock isn't enough.
*/
ret = __wt_session_get_btree(session,
dhandle->name, dhandle->checkpoint, NULL, 0);
if (ret == 0) {
ret = func(session, cfg);
if (WT_META_TRACKING(session))
WT_TRET(__wt_meta_track_handle_lock(session, 0));
else
WT_TRET(__wt_session_release_btree(session));
} else if (ret == EBUSY)
ret = __wt_conn_btree_apply_single(session, dhandle->name,
dhandle->checkpoint, func, cfg);
return (ret);
}
/*
* __lsm_discard_handle --
* Try to discard a handle from cache.
*/
static int
__lsm_discard_handle(
WT_SESSION_IMPL *session, const char *uri, const char *checkpoint)
{
WT_DECL_RET;
int locked;
/* This will fail with EBUSY if the file is still in use. */
WT_RET(__wt_session_get_btree(session, uri, checkpoint, NULL,
WT_DHANDLE_EXCLUSIVE | WT_DHANDLE_LOCK_ONLY));
WT_ASSERT(session, S2BT(session)->modified == 0);
/*
* We need the checkpoint lock to discard in-memory handles: otherwise,
* an application checkpoint could see this file locked and fail with
* EBUSY.
*
* We can't get the checkpoint lock earlier or it will deadlock with
* the schema lock.
*/
locked = 0;
if (checkpoint == NULL && (ret =
__wt_spin_trylock(session, &S2C(session)->checkpoint_lock)) == 0)
locked = 1;
if (ret == 0)
F_SET(session->dhandle, WT_DHANDLE_DISCARD);
WT_TRET(__wt_session_release_btree(session));
if (locked)
__wt_spin_unlock(session, &S2C(session)->checkpoint_lock);
return (ret);
}
/*
* __lsm_discard_handle --
* Try to discard a handle from cache.
*/
static int
__lsm_discard_handle(
WT_SESSION_IMPL *session, const char *uri, const char *checkpoint)
{
/* This will fail with EBUSY if the file is still in use. */
WT_RET(__wt_session_get_btree(session, uri, checkpoint, NULL,
WT_DHANDLE_EXCLUSIVE | WT_DHANDLE_LOCK_ONLY));
F_SET(session->dhandle, WT_DHANDLE_DISCARD_FORCE);
return (__wt_session_release_btree(session));
}
/*
* __backup_list_append --
* Append a new file name to the list, allocate space as necessary.
*/
static int
__backup_list_append(
WT_SESSION_IMPL *session, WT_CURSOR_BACKUP *cb, const char *uri)
{
WT_CURSOR_BACKUP_ENTRY *p;
WT_DATA_HANDLE *old_dhandle;
WT_DECL_RET;
const char *name;
int need_handle;
/* Leave a NULL at the end to mark the end of the list. */
WT_RET(__wt_realloc_def(session, &cb->list_allocated,
cb->list_next + 2, &cb->list));
p = &cb->list[cb->list_next];
p[0].name = p[1].name = NULL;
p[0].handle = p[1].handle = NULL;
need_handle = 0;
name = uri;
if (WT_PREFIX_MATCH(uri, "file:")) {
need_handle = 1;
name += strlen("file:");
}
/*
* !!!
* Assumes metadata file entries map one-to-one to physical files.
* To support a block manager where that's not the case, we'd need
* to call into the block manager and get a list of physical files
* that map to this logical "file". I'm not going to worry about
* that for now, that block manager might not even support physical
* copying of files by applications.
*/
WT_RET(__wt_strdup(session, name, &p->name));
/*
* If it's a file in the database, get a handle for the underlying
* object (this handle blocks schema level operations, for example
* WT_SESSION.drop or an LSM file discard after level merging).
*/
if (need_handle) {
old_dhandle = session->dhandle;
if ((ret =
__wt_session_get_btree(session, uri, NULL, NULL, 0)) == 0)
p->handle = session->dhandle;
session->dhandle = old_dhandle;
WT_RET(ret);
}
++cb->list_next;
return (0);
}
/*
* __wt_metadata_open --
* Opens the metadata file, sets session->metafile.
*/
int
__wt_metadata_open(WT_SESSION_IMPL *session)
{
if (session->metafile != NULL)
return (0);
WT_RET(__wt_session_get_btree(session, WT_METADATA_URI, NULL, NULL, 0));
session->metafile = S2BT(session);
WT_ASSERT(session, session->metafile != NULL);
/* The metafile doesn't need to stay locked -- release it. */
return (__wt_session_release_btree(session));
}
/*
* __wt_meta_btree_apply --
* Apply a function to all files listed in the metadata, apart from the
* metadata file.
*/
int
__wt_meta_btree_apply(WT_SESSION_IMPL *session,
int (*func)(WT_SESSION_IMPL *, const char *[]), const char *cfg[])
{
WT_CURSOR *cursor;
WT_DATA_HANDLE *saved_dhandle;
WT_DECL_RET;
const char *uri;
int cmp, tret;
saved_dhandle = session->dhandle;
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, "file:");
if ((tret = cursor->search_near(cursor, &cmp)) == 0 && cmp < 0)
tret = cursor->next(cursor);
for (; tret == 0; tret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor, &uri));
if (!WT_PREFIX_MATCH(uri, "file:"))
break;
else if (strcmp(uri, WT_METAFILE_URI) == 0)
continue;
/*
* We need to pull the handle into the session handle cache
* and make sure it's referenced to stop other internal code
* dropping the handle (e.g in LSM when cleaning up obsolete
* chunks). Holding the metadata lock isn't enough.
*/
ret = __wt_session_get_btree(session, uri, NULL, NULL, 0);
if (ret == 0) {
WT_SAVE_DHANDLE(session,
ret = func(session, cfg));
if (WT_META_TRACKING(session))
WT_TRET(
__wt_meta_track_handle_lock(session, 0));
else
WT_TRET(__wt_session_release_btree(session));
} else if (ret == EBUSY)
ret = __wt_conn_btree_apply_single(
session, uri, NULL, func, cfg);
WT_ERR(ret);
}
if (tret != WT_NOTFOUND)
WT_TRET(tret);
err: WT_TRET(cursor->close(cursor));
session->dhandle = saved_dhandle;
return (ret);
}
/*
* __drop_file --
* Drop a file.
*/
static int
__drop_file(
WT_SESSION_IMPL *session, const char *uri, int force, const char *cfg[])
{
WT_DECL_RET;
int exist;
const char *filename;
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
return (EINVAL);
if (session->btree == NULL &&
(ret = __wt_session_get_btree(session, uri, NULL, cfg,
WT_BTREE_EXCLUSIVE | WT_BTREE_LOCK_ONLY)) != 0) {
if (ret == WT_NOTFOUND || ret == ENOENT)
ret = 0;
return (ret);
}
/* Close all btree handles associated with this file. */
WT_RET(__wt_conn_btree_close_all(session, uri));
/* Remove the metadata entry (ignore missing items). */
WT_TRET(__wt_metadata_remove(session, uri));
if (force && ret == WT_NOTFOUND)
ret = 0;
/* Remove the underlying physical file. */
exist = 0;
WT_TRET(__wt_exist(session, filename, &exist));
if (exist) {
/*
* There is no point tracking this operation: there is no going
* back from here.
*/
WT_TRET(__wt_remove(session, filename));
}
return (ret);
}
/*
* __meta_btree_apply --
* Apply a function to all files listed in the metadata, apart from the
* metadata file.
*/
static inline int
__meta_btree_apply(WT_SESSION_IMPL *session, WT_CURSOR *cursor,
int (*file_func)(WT_SESSION_IMPL *, const char *[]),
int (*name_func)(WT_SESSION_IMPL *, const char *, bool *),
const char *cfg[])
{
WT_DECL_RET;
const char *uri;
bool skip;
while ((ret = cursor->next(cursor)) == 0) {
WT_RET(cursor->get_key(cursor, &uri));
if (strcmp(uri, WT_METAFILE_URI) == 0)
continue;
skip = false;
if (name_func != NULL)
WT_RET(name_func(session, uri, &skip));
if (file_func == NULL || skip || !WT_PREFIX_MATCH(uri, "file:"))
continue;
/*
* We need to pull the handle into the session handle cache
* and make sure it's referenced to stop other internal code
* dropping the handle (e.g in LSM when cleaning up obsolete
* chunks). Holding the metadata lock isn't enough.
*/
if ((ret = __wt_session_get_btree(
session, uri, NULL, NULL, 0)) != 0)
return (ret == EBUSY ? 0 : ret);
WT_SAVE_DHANDLE(session, ret = file_func(session, cfg));
WT_TRET(__wt_session_release_btree(session));
WT_RET(ret);
}
WT_RET_NOTFOUND_OK(ret);
return (0);
}
/*
* __compact_handle_append --
* Gather a file handle to be compacted.
* Called via the schema_worker function.
*/
static int
__compact_handle_append(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_DECL_RET;
WT_UNUSED(cfg);
WT_RET(__wt_session_get_btree(
session, session->dhandle->name, NULL, NULL, 0));
/* Set compact active on the handle. */
if ((ret = __compact_start(session)) != 0) {
WT_TRET(__wt_session_release_btree(session));
return (ret);
}
/* Make sure there is space for the next entry. */
WT_RET(__wt_realloc_def(session, &session->op_handle_allocated,
session->op_handle_next + 1, &session->op_handle));
session->op_handle[session->op_handle_next++] = session->dhandle;
return (0);
}
/*
* __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);
}
/*
* __create_file --
* Create a new 'file:' object.
*/
static int
__create_file(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_DECL_ITEM(val);
WT_DECL_RET;
uint32_t allocsize;
int is_metadata;
const char *filename, **p, *filecfg[] =
{ WT_CONFIG_BASE(session, file_meta), config, NULL, NULL };
char *fileconf;
fileconf = NULL;
is_metadata = strcmp(uri, WT_METAFILE_URI) == 0;
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
WT_RET_MSG(session, EINVAL, "Expected a 'file:' URI: %s", uri);
/* Check if the file already exists. */
if (!is_metadata && (ret =
__wt_metadata_search(session, uri, &fileconf)) != WT_NOTFOUND) {
if (exclusive)
WT_TRET(EEXIST);
goto err;
}
/* Sanity check the allocation size. */
WT_RET(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
/* Create the file. */
WT_ERR(__wt_block_manager_create(session, filename, allocsize));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, NULL, uri));
/*
* If creating an ordinary file, append the file ID and current version
* numbers to the passed-in configuration and insert the resulting
* configuration into the metadata.
*/
if (!is_metadata) {
WT_ERR(__wt_scr_alloc(session, 0, &val));
WT_ERR(__wt_buf_fmt(session, val,
"id=%" PRIu32 ",version=(major=%d,minor=%d)",
++S2C(session)->next_file_id,
WT_BTREE_MAJOR_VERSION_MAX, WT_BTREE_MINOR_VERSION_MAX));
for (p = filecfg; *p != NULL; ++p)
;
*p = val->data;
WT_ERR(__wt_config_collapse(session, filecfg, &fileconf));
WT_ERR(__wt_metadata_insert(session, uri, fileconf));
}
/*
* Open the file to check that it was setup correctly. We don't need to
* pass the configuration, we just wrote the collapsed configuration
* into the metadata file, and it's going to be read/used by underlying
* functions.
*
* Keep the handle exclusive until it is released at the end of the
* call, otherwise we could race with a drop.
*/
WT_ERR(__wt_session_get_btree(
session, uri, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_handle_lock(session, 1));
else
WT_ERR(__wt_session_release_btree(session));
err: __wt_scr_free(session, &val);
__wt_free(session, fileconf);
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);
}
/*
* __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);
}