/*
* __wt_metadata_update --
* Update a row in the metadata.
*/
int
__wt_metadata_update(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Update: key: %s, value: %s, tracking: %s, %s" "turtle",
key, value, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key))
return (__wt_turtle_update(session, key, value));
if (WT_META_TRACKING(session))
WT_RET(__wt_meta_track_update(session, key));
WT_RET(__wt_metadata_cursor(session, "overwrite", &cursor));
cursor->set_key(cursor, key);
cursor->set_value(cursor, value);
WT_ERR(cursor->insert(cursor));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_metadata_insert --
* Insert a row into the metadata.
*/
int
__wt_metadata_insert(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Insert: key: %s, value: %s, tracking: %s, %s" "turtle",
key, value, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key))
WT_RET_MSG(session, EINVAL,
"%s: insert not supported on the turtle file", key);
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, key);
cursor->set_value(cursor, value);
WT_ERR(cursor->insert(cursor));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_insert(session, key));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_metadata_update --
* Update a row in the metadata.
*/
int
__wt_metadata_update(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Update: key: %s, value: %s, tracking: %s, %s" "turtle",
key, value, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key)) {
WT_WITH_TURTLE_LOCK(session, ret,
ret = __wt_turtle_update(session, key, value));
return (ret);
}
if (WT_META_TRACKING(session))
WT_RET(__wt_meta_track_update(session, key));
WT_RET(__wt_metadata_cursor(session, &cursor));
/* This cursor needs to have overwrite semantics. */
WT_ASSERT(session, F_ISSET(cursor, WT_CURSTD_OVERWRITE));
cursor->set_key(cursor, key);
cursor->set_value(cursor, value);
WT_ERR(cursor->insert(cursor));
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
return (ret);
}
/*
* __wt_metadata_remove --
* Remove a row from the metadata.
*/
int
__wt_metadata_remove(WT_SESSION_IMPL *session, const char *key)
{
WT_CURSOR *cursor;
WT_DECL_RET;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Remove: key: %s, tracking: %s, %s" "turtle",
key, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key))
WT_RET_MSG(session, EINVAL,
"%s: remove not supported on the turtle file", key);
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, key);
WT_ERR(cursor->search(cursor));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_update(session, key));
WT_ERR(cursor->remove(cursor));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __rename_file --
* WT_SESSION::rename for a file.
*/
static int
__rename_file(
WT_SESSION_IMPL *session, const char *uri, const char *newuri)
{
WT_DECL_RET;
bool exist;
const char *filename, *newfile;
char *newvalue, *oldvalue;
newvalue = oldvalue = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") ||
!WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/* Close any btree handles in the file. */
WT_WITH_HANDLE_LIST_LOCK(session,
ret = __wt_conn_dhandle_close_all(session, uri, false));
WT_ERR(ret);
/*
* First, check if the file being renamed exists in the system. Doing
* this check first matches the table rename behavior because we return
* WT_NOTFOUND when the renamed file doesn't exist (subsequently mapped
* to ENOENT by the session layer).
*/
WT_ERR(__wt_metadata_search(session, uri, &oldvalue));
/*
* Check to see if the proposed name is already in use, in either the
* metadata or the filesystem.
*/
switch (ret = __wt_metadata_search(session, newuri, &newvalue)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
/* NOTREACHED */
case WT_NOTFOUND:
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_fs_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, oldvalue));
/* Rename the underlying file. */
WT_ERR(__wt_fs_rename(session, filename, newfile, false));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err: __wt_free(session, newvalue);
__wt_free(session, oldvalue);
return (ret);
}
/*
* __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);
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, int unroll)
{
WT_BTREE *saved_btree;
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
const char *ckpt_cfg[] = API_CONF_DEFAULTS(session, checkpoint, NULL);
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
if (--session->meta_track_nest != 0)
return (0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
while (--trk >= trk_orig)
WT_TRET(__meta_track_apply(session, trk, unroll));
/* If the operation succeeded, checkpoint the metadata. */
if (!unroll && ret == 0 && session->metafile != NULL) {
saved_btree = session->btree;
session->btree = session->metafile;
ret = __wt_checkpoint(session, ckpt_cfg);
session->btree = saved_btree;
}
return (ret);
}
/*
* __wt_metadata_search --
* Return a copied row from the metadata.
* The caller is responsible for freeing the allocated memory.
*/
int
__wt_metadata_search(
WT_SESSION_IMPL *session, const char *key, char **valuep)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *value;
*valuep = NULL;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Search: key: %s, tracking: %s, %s" "turtle",
key, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key))
return (__wt_turtle_read(session, key, valuep));
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, key);
WT_ERR(cursor->search(cursor));
WT_ERR(cursor->get_value(cursor, &value));
WT_ERR(__wt_strdup(session, value, valuep));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*修改文件名操作*/
static int __rename_file(WT_SESSION_IMPL* session, const char* uri, const char* newuri)
{
WT_DECL_RET;
int exist;
const char *filename, *newfile;
char *newvalue, *oldvalue;
newvalue = oldvalue = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") || !WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/*关闭这个文件对应的data source handler*/
WT_WITH_DHANDLE_LOCK(session, ret = __wt_conn_dhandle_close_all(session, uri, 0));
WT_ERR(ret);
/*
* First, check if the file being renamed exists in the system. Doing
* this check first matches the table rename behavior because we return
* WT_NOTFOUND when the renamed file doesn't exist (subsequently mapped
* to ENOENT by the session layer).
*/
/*先查出旧的文件名uri对应的meta 信息，再用新的文件名uri旧版本的meta信息*/
WT_ERR(__wt_metadata_search(session, uri, &oldvalue));
switch (ret = __wt_metadata_search(session, newuri, &newvalue)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
/* NOTREACHED */
case WT_NOTFOUND:
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, oldvalue));
WT_ERR(__wt_rename(session, filename, newfile));
if(WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err:
__wt_free(session, newvalue);
__wt_free(session, oldvalue);
return ret;
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, int need_sync, int unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, return now and avoid unnecessary
* metadata checkpoints. For example, this happens if attempting to
* create a data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
return (0);
while (--trk >= trk_orig)
WT_TRET(__meta_track_apply(session, trk, unroll));
/*
* Unroll operations don't need to flush the metadata.
*
* Also, if we don't have the metadata handle (e.g, we're in the
* process of creating the metadata), we can't sync it.
*/
if (unroll || ret != 0 || !need_sync || session->meta_dhandle == NULL)
return (ret);
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED)) {
if (!FLD_ISSET(S2C(session)->txn_logsync,
WT_LOG_DSYNC | WT_LOG_FSYNC))
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_txn_checkpoint_log(session,
0, WT_TXN_LOG_CKPT_SYNC, NULL));
} else {
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_checkpoint(session, NULL));
WT_RET(ret);
WT_WITH_DHANDLE(session, session->meta_dhandle,
ret = __wt_checkpoint_sync(session, NULL));
}
return (ret);
}
/*
* __rename_file --
* WT_SESSION::rename for a file.
*/
static int
__rename_file(
WT_SESSION_IMPL *session, const char *uri, const char *newuri)
{
WT_DECL_RET;
int exist;
const char *filename, *newfile, *value;
value = NULL;
filename = uri;
newfile = newuri;
if (!WT_PREFIX_SKIP(filename, "file:") ||
!WT_PREFIX_SKIP(newfile, "file:"))
return (EINVAL);
/* Close any btree handles in the file. */
WT_RET(__wt_conn_btree_close_all(session, uri));
/*
* Check to see if the proposed name is already in use, in either
* the metadata or the filesystem.
*/
switch (ret = __wt_metadata_read(session, newuri, &value)) {
case 0:
WT_ERR_MSG(session, EEXIST, "%s", newuri);
case WT_NOTFOUND:
ret = 0;
break;
default:
WT_ERR(ret);
}
WT_ERR(__wt_exist(session, newfile, &exist));
if (exist)
WT_ERR_MSG(session, EEXIST, "%s", newfile);
/* Replace the old file entries with new file entries. */
WT_ERR(__wt_metadata_read(session, uri, &value));
WT_ERR(__wt_metadata_remove(session, uri));
WT_ERR(__wt_metadata_insert(session, newuri, value));
/* Rename the underlying file. */
WT_ERR(__wt_rename(session, filename, newfile));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, uri, newuri));
err: __wt_free(session, value);
return (ret);
}
/*
* __wt_metadata_search --
* Return a copied row from the metadata.
* The caller is responsible for freeing the allocated memory.
*/
int
__wt_metadata_search(WT_SESSION_IMPL *session, const char *key, char **valuep)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *value;
*valuep = NULL;
__wt_verbose(session, WT_VERB_METADATA,
"Search: key: %s, tracking: %s, %s" "turtle",
key, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not ");
if (__metadata_turtle(key)) {
/*
* The returned value should only be set if ret is non-zero, but
* Coverity is convinced otherwise. The code path is used enough
* that Coverity complains a lot, add an error check to get some
* peace and quiet.
*/
if ((ret = __wt_turtle_read(session, key, valuep)) != 0)
__wt_free(session, *valuep);
return (ret);
}
/*
* All metadata reads are at read-uncommitted isolation. That's
* because once a schema-level operation completes, subsequent
* operations must see the current version of checkpoint metadata, or
* they may try to read blocks that may have been freed from a file.
* Metadata updates use non-transactional techniques (such as the
* schema and metadata locks) to protect access to in-flight updates.
*/
WT_RET(__wt_metadata_cursor(session, &cursor));
cursor->set_key(cursor, key);
WT_WITH_TXN_ISOLATION(session, WT_ISO_READ_UNCOMMITTED,
ret = cursor->search(cursor));
WT_ERR(ret);
WT_ERR(cursor->get_value(cursor, &value));
WT_ERR(__wt_strdup(session, value, valuep));
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
if (ret != 0)
__wt_free(session, *valuep);
return (ret);
}
/*
* __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);
}
/*
* __wt_metadata_remove --
* Remove a row from the metadata.
*/
int
__wt_metadata_remove(WT_SESSION_IMPL *session, const char *key)
{
WT_CURSOR *cursor;
WT_DECL_RET;
if (__metadata_turtle(key))
WT_RET_MSG(session, EINVAL,
"%s: remove not supported on the turtle file", key);
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, key);
WT_ERR(cursor->search(cursor));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_update(session, key));
WT_ERR(cursor->remove(cursor));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_metadata_insert --
* Insert a row into the metadata.
*/
int
__wt_metadata_insert(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_CURSOR *cursor;
WT_DECL_RET;
if (__metadata_turtle(key))
WT_RET_MSG(session, EINVAL,
"%s: insert not supported on the turtle file", key);
WT_RET(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, key);
cursor->set_value(cursor, value);
WT_ERR(cursor->insert(cursor));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_insert(session, key));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_metadata_update --
* Update a row in the metadata.
*/
int
__wt_metadata_update(
WT_SESSION_IMPL *session, const char *key, const char *value)
{
WT_CURSOR *cursor;
WT_DECL_RET;
if (__metadata_turtle(key))
return (__wt_turtle_update(session, key, value));
if (WT_META_TRACKING(session))
WT_RET(__wt_meta_track_update(session, key));
WT_RET(__wt_metadata_cursor(session, "overwrite", &cursor));
cursor->set_key(cursor, key);
cursor->set_value(cursor, value);
WT_ERR(cursor->insert(cursor));
err: WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_meta_track_sub_off --
* Commit a group of operations independent of the main transaction.
*/
int
__wt_meta_track_sub_off(WT_SESSION_IMPL *session)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
if (!WT_META_TRACKING(session) || session->meta_track_sub == NULL)
return (0);
trk_orig = session->meta_track_sub;
trk = session->meta_track_next;
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
while (--trk >= trk_orig)
WT_TRET(__meta_track_apply(session, trk));
session->meta_track_next = trk_orig;
return (ret);
}
/*
* __wt_meta_track_find_handle --
* Check if we have already seen a handle.
*/
int
__wt_meta_track_find_handle(
WT_SESSION_IMPL *session, const char *name, const char *checkpoint)
{
WT_META_TRACK *trk, *trk_orig;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
while (--trk >= trk_orig) {
if (trk->op != WT_ST_LOCK)
continue;
if (strcmp(trk->dhandle->name, name) == 0 &&
((trk->dhandle->checkpoint == NULL && checkpoint == NULL) ||
(trk->dhandle->checkpoint != NULL &&
strcmp(trk->dhandle->checkpoint, checkpoint) == 0)))
return (0);
}
return (WT_NOTFOUND);
}
/*
* __wt_metadata_search --
* Return a copied row from the metadata.
* The caller is responsible for freeing the allocated memory.
*/
int
__wt_metadata_search(WT_SESSION_IMPL *session, const char *key, char **valuep)
{
WT_CURSOR *cursor;
WT_DECL_RET;
const char *value;
*valuep = NULL;
WT_RET(__wt_verbose(session, WT_VERB_METADATA,
"Search: key: %s, tracking: %s, %s" "turtle",
key, WT_META_TRACKING(session) ? "true" : "false",
__metadata_turtle(key) ? "" : "not "));
if (__metadata_turtle(key))
return (__wt_turtle_read(session, key, valuep));
/*
* All metadata reads are at read-uncommitted isolation. That's
* because once a schema-level operation completes, subsequent
* operations must see the current version of checkpoint metadata, or
* they may try to read blocks that may have been freed from a file.
* Metadata updates use non-transactional techniques (such as the
* schema and metadata locks) to protect access to in-flight updates.
*/
WT_RET(__wt_metadata_cursor(session, &cursor));
cursor->set_key(cursor, key);
WT_WITH_TXN_ISOLATION(session, WT_ISO_READ_UNCOMMITTED,
ret = cursor->search(cursor));
WT_ERR(ret);
WT_ERR(cursor->get_value(cursor, &value));
WT_ERR(__wt_strdup(session, value, valuep));
err: WT_TRET(__wt_metadata_cursor_release(session, &cursor));
return (ret);
}
/*
* __snapshot_worker --
* Snapshot the tree.
*/
static int
__snapshot_worker(
WT_SESSION_IMPL *session, const char *name, int discard, snapshot_op op)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_SNAPSHOT *deleted, *snap, *snapbase;
int force, matched, tracked;
btree = session->btree;
matched = tracked = 0;
snap = snapbase = NULL;
/* Snapshots are single-threaded. */
__wt_writelock(session, btree->snaplock);
/* Set the name to the default, if we aren't provided one. */
if (op == SNAPSHOT && name == NULL) {
force = 0;
name = WT_INTERNAL_SNAPSHOT;
} else
force = 1;
/*
* Get the list of snapshots for this file. If there's no reference,
* this file is dead. Discard it from the cache without bothering to
* write any dirty pages.
*/
if ((ret =
__wt_meta_snaplist_get(session, btree->name, &snapbase)) != 0) {
if (ret == WT_NOTFOUND)
ret = __wt_bt_cache_flush(
session, NULL, WT_SYNC_DISCARD_NOWRITE, 0);
goto err;
}
switch (op) {
case SNAPSHOT:
/*
* Create a new, possibly named, snapshot. Review existing
* snapshots, deleting default snapshots and snapshots with
* matching names, add the new snapshot entry at the end of
* the list.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap)
if (strcmp(snap->name, name) == 0 ||
strcmp(snap->name, WT_INTERNAL_SNAPSHOT) == 0)
F_SET(snap, WT_SNAP_DELETE);
WT_ERR(__wt_strdup(session, name, &snap->name));
F_SET(snap, WT_SNAP_ADD);
break;
case SNAPSHOT_DROP:
/*
* Drop all snapshots with matching names.
* Drop all snapshots with the default name.
* Add a new snapshot with the default name.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap) {
/*
* There should be only one snapshot with a matching
* name, but it doesn't hurt to check the rest.
*/
if (strcmp(snap->name, name) == 0)
matched = 1;
else if (strcmp(snap->name, WT_INTERNAL_SNAPSHOT) != 0)
continue;
F_SET(snap, WT_SNAP_DELETE);
}
if (!matched)
goto nomatch;
WT_ERR(__wt_strdup(session, WT_INTERNAL_SNAPSHOT, &snap->name));
F_SET(snap, WT_SNAP_ADD);
break;
case SNAPSHOT_DROP_ALL:
/*
* Drop all snapshots.
* Add a new snapshot with the default name.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap)
F_SET(snap, WT_SNAP_DELETE);
WT_ERR(__wt_strdup(session, WT_INTERNAL_SNAPSHOT, &snap->name));
F_SET(snap, WT_SNAP_ADD);
break;
case SNAPSHOT_DROP_FROM:
/*
* Drop all snapshots after, and including, the named snapshot.
* Drop all snapshots with the default name.
* Add a new snapshot with the default name.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap) {
if (strcmp(snap->name, name) == 0)
matched = 1;
if (matched ||
strcmp(snap->name, WT_INTERNAL_SNAPSHOT) == 0)
F_SET(snap, WT_SNAP_DELETE);
}
if (!matched)
goto nomatch;
WT_ERR(__wt_strdup(session, WT_INTERNAL_SNAPSHOT, &snap->name));
F_SET(snap, WT_SNAP_ADD);
break;
case SNAPSHOT_DROP_TO:
/*
* Drop all snapshots before, and including, the named snapshot.
* Drop all snapshots with the default name.
* Add a new snapshot with the default name.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap) {
if (!matched ||
strcmp(snap->name, WT_INTERNAL_SNAPSHOT) == 0)
F_SET(snap, WT_SNAP_DELETE);
if (strcmp(snap->name, name) == 0)
matched = 1;
}
if (!matched)
nomatch: WT_ERR_MSG(session,
EINVAL, "no snapshot named %s was found", name);
WT_ERR(__wt_strdup(session, WT_INTERNAL_SNAPSHOT, &snap->name));
F_SET(snap, WT_SNAP_ADD);
break;
}
/*
* Lock the snapshots that will be deleted.
*
* Snapshots are only locked when tracking is enabled, which covers
* sync and drop operations, but not close. The reasoning is that
* there should be no access to a snapshot during close, because any
* thread accessing a snapshot will also have the current file handle
* open.
*/
if (WT_META_TRACKING(session))
WT_SNAPSHOT_FOREACH(snapbase, deleted)
if (F_ISSET(deleted, WT_SNAP_DELETE))
WT_ERR(__wt_session_lock_snapshot(session,
deleted->name, WT_BTREE_EXCLUSIVE));
WT_ERR(__wt_bt_cache_flush(
session, snapbase, discard ? WT_SYNC_DISCARD : WT_SYNC, force));
/* If there was a snapshot, update the metadata. */
if (snap->raw.data == NULL) {
if (force)
WT_ERR_MSG(session,
EINVAL, "cache flush failed to create a snapshot");
} else {
WT_ERR(__wt_meta_snaplist_set(session, btree->name, snapbase));
/*
* If tracking is enabled, defer making pages available until
* the end of the transaction. The exception is if the handle
* is being discarded: in that case, it will be gone by the
* time we try to apply or unroll the meta tracking event.
*/
if (WT_META_TRACKING(session) && !discard) {
WT_ERR(__wt_meta_track_checkpoint(session));
tracked = 1;
} else
WT_ERR(__wt_bm_snapshot_resolve(session, snapbase));
}
err: __wt_meta_snaplist_free(session, snapbase);
if (!tracked)
__wt_rwunlock(session, btree->snaplock);
return (ret);
}
/*
* __wt_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_SESSION_IMPL *ckpt_session;
int saved_ret;
bool did_drop;
saved_ret = 0;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, skip unnecessary metadata
* checkpoints. For example, this happens if attempting to create a
* data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
goto err;
/* Unrolling doesn't require syncing the metadata. */
if (unroll)
goto err;
if (F_ISSET(session, WT_SESSION_SCHEMA_TXN)) {
F_CLR(session, WT_SESSION_SCHEMA_TXN);
#ifdef WT_ENABLE_SCHEMA_TXN
WT_ERR(__wt_txn_commit(session, NULL));
__wt_errx(session, "TRACK: Commit internal schema txn");
#endif
}
/*
* If we don't have the metadata cursor (e.g, we're in the process of
* creating the metadata), we can't sync it.
*/
if (!need_sync || session->meta_cursor == NULL ||
F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto err;
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED))
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_txn_checkpoint_log(
session, false, WT_TXN_LOG_CKPT_SYNC, NULL));
else {
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
ckpt_session = S2C(session)->meta_ckpt_session;
/*
* If this operation is part of a running transaction, that
* should be included in the checkpoint.
*/
ckpt_session->txn.id = session->txn.id;
WT_ASSERT(session,
!F_ISSET(session, WT_SESSION_LOCKED_METADATA));
WT_WITH_DHANDLE(ckpt_session, WT_SESSION_META_DHANDLE(session),
WT_WITH_METADATA_LOCK(ckpt_session,
ret = __wt_checkpoint(ckpt_session, NULL)));
ckpt_session->txn.id = WT_TXN_NONE;
if (ret == 0)
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint_sync(session, NULL));
}
err: /*
* Undo any tracked operations on failure.
* Apply any tracked operations post-commit.
*/
did_drop = false;
if (unroll || ret != 0) {
saved_ret = ret;
ret = 0;
while (--trk >= trk_orig) {
did_drop = did_drop || trk->op == WT_ST_DROP_COMMIT;
WT_TRET(__meta_track_unroll(session, trk));
}
} else
for (; trk_orig < trk; trk_orig++) {
did_drop = did_drop ||
trk_orig->op == WT_ST_DROP_COMMIT;
WT_TRET(__meta_track_apply(session, trk_orig));
}
if (F_ISSET(session, WT_SESSION_SCHEMA_TXN)) {
F_CLR(session, WT_SESSION_SCHEMA_TXN);
/*
* We should have committed above unless we're unrolling, there
* was an error or the operation was a noop.
*/
WT_ASSERT(session, unroll || saved_ret != 0 ||
session->txn.mod_count == 0);
#ifdef WT_ENABLE_SCHEMA_TXN
__wt_err(session, saved_ret,
"TRACK: Abort internal schema txn");
WT_TRET(__wt_txn_rollback(session, NULL));
#endif
}
/*
* Wake up the sweep thread: particularly for the in-memory
* storage engine, we want to reclaim space immediately.
*/
if (did_drop && S2C(session)->sweep_cond != NULL)
__wt_cond_signal(session, S2C(session)->sweep_cond);
if (ret != 0)
WT_PANIC_RET(session, ret,
"failed to apply or unroll all tracked operations");
return (saved_ret == 0 ? 0 : saved_ret);
}
/*
* __drop_table --
* WT_SESSION::drop for a table.
*/
static int
__drop_table(
WT_SESSION_IMPL *session, const char *uri, const char *cfg[])
{
WT_COLGROUP *colgroup;
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
u_int i;
const char *name;
bool tracked;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_TABLE_WRITE));
name = uri;
WT_PREFIX_SKIP_REQUIRED(session, name, "table:");
table = NULL;
tracked = false;
/*
* Open the table so we can drop its column groups and indexes.
*
* Ideally we would keep the table locked exclusive across the drop,
* but for now we rely on the global table lock to prevent the table
* being reopened while it is being dropped. One issue is that the
* WT_WITHOUT_LOCKS macro can drop and reacquire the global table lock,
* avoiding deadlocks while waiting for LSM operation to quiesce.
*
* Temporarily getting the table exclusively serves the purpose
* of ensuring that cursors on the table that are already open
* must at least be closed before this call proceeds.
*/
WT_ERR(__wt_schema_get_table_uri(session, uri, true,
WT_DHANDLE_EXCLUSIVE, &table));
WT_ERR(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(session, uri, true, 0, &table));
/* Drop the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if ((colgroup = table->cgroups[i]) == NULL)
continue;
/*
* Drop the column group before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, colgroup->source, cfg));
WT_ERR(__wt_metadata_remove(session, colgroup->name));
}
/* Drop the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
if ((idx = table->indices[i]) == NULL)
continue;
/*
* Drop the index before updating the metadata to avoid
* the metadata for the table becoming inconsistent if we can't
* get exclusive access.
*/
WT_ERR(__wt_schema_drop(session, idx->source, cfg));
WT_ERR(__wt_metadata_remove(session, idx->name));
}
/* Make sure the table data handle is closed. */
WT_TRET(__wt_schema_release_table(session, table));
WT_ERR(__wt_schema_get_table_uri(
session, uri, true, WT_DHANDLE_EXCLUSIVE, &table));
F_SET(&table->iface, WT_DHANDLE_DISCARD);
if (WT_META_TRACKING(session)) {
WT_WITH_DHANDLE(session, &table->iface,
ret = __wt_meta_track_handle_lock(session, false));
WT_ERR(ret);
tracked = true;
}
/* Remove the metadata entry (ignore missing items). */
WT_ERR(__wt_metadata_remove(session, uri));
err: if (table != NULL && !tracked)
WT_TRET(__wt_schema_release_table(session, table));
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_meta_track_off --
* Turn off metadata operation tracking, unrolling on error.
*/
int
__wt_meta_track_off(WT_SESSION_IMPL *session, bool need_sync, bool unroll)
{
WT_DECL_RET;
WT_META_TRACK *trk, *trk_orig;
WT_SESSION_IMPL *ckpt_session;
WT_ASSERT(session,
WT_META_TRACKING(session) && session->meta_track_nest > 0);
trk_orig = session->meta_track;
trk = session->meta_track_next;
/* If it was a nested transaction, there is nothing to do. */
if (--session->meta_track_nest != 0)
return (0);
/* Turn off tracking for unroll. */
session->meta_track_next = session->meta_track_sub = NULL;
/*
* If there were no operations logged, return now and avoid unnecessary
* metadata checkpoints. For example, this happens if attempting to
* create a data source that already exists (or drop one that doesn't).
*/
if (trk == trk_orig)
return (0);
if (unroll) {
while (--trk >= trk_orig)
WT_TRET(__meta_track_unroll(session, trk));
/* Unroll operations don't need to flush the metadata. */
return (ret);
}
/*
* If we don't have the metadata cursor (e.g, we're in the process of
* creating the metadata), we can't sync it.
*/
if (!need_sync || session->meta_cursor == NULL ||
F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
goto done;
/* If we're logging, make sure the metadata update was flushed. */
if (FLD_ISSET(S2C(session)->log_flags, WT_CONN_LOG_ENABLED)) {
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_txn_checkpoint_log(
session, false, WT_TXN_LOG_CKPT_SYNC, NULL));
WT_RET(ret);
} else {
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SCHEMA));
ckpt_session = S2C(session)->meta_ckpt_session;
/*
* If this operation is part of a running transaction, that
* should be included in the checkpoint.
*/
ckpt_session->txn.id = session->txn.id;
F_SET(ckpt_session, WT_SESSION_LOCKED_METADATA);
WT_WITH_METADATA_LOCK(session, ret,
WT_WITH_DHANDLE(ckpt_session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint(ckpt_session, NULL)));
F_CLR(ckpt_session, WT_SESSION_LOCKED_METADATA);
ckpt_session->txn.id = WT_TXN_NONE;
WT_RET(ret);
WT_WITH_DHANDLE(session,
WT_SESSION_META_DHANDLE(session),
ret = __wt_checkpoint_sync(session, NULL));
WT_RET(ret);
}
done: /* Apply any tracked operations post-commit. */
for (; trk_orig < trk; trk_orig++)
WT_TRET(__meta_track_apply(session, trk_orig));
return (ret);
}