diff options
Diffstat (limited to 'src/lsm')
| -rw-r--r-- | src/lsm/lsm_cursor.c | 116 | ||||
| -rw-r--r-- | src/lsm/lsm_manager.c | 66 | ||||
| -rw-r--r-- | src/lsm/lsm_merge.c | 4 | ||||
| -rw-r--r-- | src/lsm/lsm_meta.c | 2 | ||||
| -rw-r--r-- | src/lsm/lsm_stat.c | 12 | ||||
| -rw-r--r-- | src/lsm/lsm_tree.c | 73 | ||||
| -rw-r--r-- | src/lsm/lsm_work_unit.c | 62 | ||||
| -rw-r--r-- | src/lsm/lsm_worker.c | 25 |
8 files changed, 199 insertions, 161 deletions
diff --git a/src/lsm/lsm_cursor.c b/src/lsm/lsm_cursor.c index 839648b97d7..52265f02e62 100644 --- a/src/lsm/lsm_cursor.c +++ b/src/lsm/lsm_cursor.c @@ -10,10 +10,10 @@ #define WT_FORALL_CURSORS(clsm, c, i) \ for ((i) = (clsm)->nchunks; (i) > 0;) \ - if (((c) = (clsm)->chunks[--i]->cursor) != NULL) + if (((c) = (clsm)->chunks[--(i)]->cursor) != NULL) #define WT_LSM_CURCMP(s, lsm_tree, c1, c2, cmp) \ - __wt_compare(s, (lsm_tree)->collator, &(c1)->key, &(c2)->key, &cmp) + __wt_compare(s, (lsm_tree)->collator, &(c1)->key, &(c2)->key, &(cmp)) static int __clsm_lookup(WT_CURSOR_LSM *, WT_ITEM *); static int __clsm_open_cursors(WT_CURSOR_LSM *, bool, u_int, uint32_t); @@ -178,20 +178,12 @@ __clsm_enter(WT_CURSOR_LSM *clsm, bool reset, bool update) if (reset) { WT_ASSERT(session, !F_ISSET(&clsm->iface, - WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT)); + WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT)); WT_RET(__clsm_reset_cursors(clsm, NULL)); } for (;;) { - /* - * If the cursor looks up-to-date, check if the cache is full. - * In case this call blocks, the check will be repeated before - * proceeding. - */ - if (clsm->dsk_gen != lsm_tree->dsk_gen && - lsm_tree->nchunks != 0) - goto open; - + /* Check if the cursor looks up-to-date. */ if (clsm->dsk_gen != lsm_tree->dsk_gen && lsm_tree->nchunks != 0) goto open; @@ -304,7 +296,7 @@ __clsm_leave(WT_CURSOR_LSM *clsm) * byte, if the application uses two leading DC4 byte for some reason, we'll do * a wasted data copy each time a new value is inserted into the object. */ -static const WT_ITEM __tombstone = { "\x14\x14", 2, 0, NULL, 0 }; +static const WT_ITEM __tombstone = { "\x14\x14", 2, NULL, 0, 0 }; /* * __clsm_deleted -- @@ -666,7 +658,7 @@ retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) { */ if (i != nchunks - 1) clsm->chunks[i]->cursor->insert = - __wt_curfile_update_check; + __wt_curfile_insert_check; if (!F_ISSET(clsm, WT_CLSM_MERGE) && F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) @@ -688,19 +680,29 @@ retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) { if (chunk != NULL && !F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && chunk->switch_txn == WT_TXN_NONE) { - clsm->primary_chunk = chunk; primary = clsm->chunks[clsm->nchunks - 1]->cursor; + btree = ((WT_CURSOR_BTREE *)primary)->btree; + /* - * Disable eviction for the in-memory chunk. Also clear the - * bulk load flag here, otherwise eviction will be enabled by - * the first update. + * If the primary is not yet set as the primary, do that now. + * Note that eviction was configured off when the underlying + * object was created, which is what we want, leave it alone. + * + * We don't have to worry about races here: every thread that + * modifies the tree will have to come through here, at worse + * we set the flag repeatedly. We don't use a WT_BTREE handle + * flag, however, we could race doing the read-modify-write of + * the flags field. + * + * If something caused the chunk to be closed and reopened + * since it was created, we can no longer use it as a primary + * chunk and we need to force a switch. We detect the tree was + * created when it was opened by checking the "original" flag. */ - btree = ((WT_CURSOR_BTREE *)(primary))->btree; - if (btree->bulk_load_ok) { - btree->bulk_load_ok = false; - WT_WITH_BTREE(session, btree, - __wt_btree_lsm_switch_primary(session, true)); - } + if (!btree->lsm_primary && btree->original) + btree->lsm_primary = true; + if (btree->lsm_primary) + clsm->primary_chunk = chunk; } clsm->dsk_gen = lsm_tree->dsk_gen; @@ -1213,7 +1215,8 @@ __clsm_lookup(WT_CURSOR_LSM *clsm, WT_ITEM *value) WT_LSM_TREE_STAT_INCR( session, clsm->lsm_tree->bloom_miss); continue; - } else if (ret == 0) + } + if (ret == 0) WT_LSM_TREE_STAT_INCR( session, clsm->lsm_tree->bloom_hit); WT_ERR(ret); @@ -1239,10 +1242,10 @@ __clsm_lookup(WT_CURSOR_LSM *clsm, WT_ITEM *value) WT_ERR(WT_NOTFOUND); done: -err: F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET); - if (ret == 0) { - clsm->current = c; +err: if (ret == 0) { + F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET); F_SET(cursor, WT_CURSTD_KEY_INT); + clsm->current = c; if (value == &cursor->value) F_SET(cursor, WT_CURSTD_VALUE_INT); } else if (c != NULL) @@ -1318,7 +1321,8 @@ __clsm_search_near(WT_CURSOR *cursor, int *exactp) if ((ret = c->search_near(c, &cmp)) == WT_NOTFOUND) { ret = 0; continue; - } else if (ret != 0) + } + if (ret != 0) goto err; /* Do we have an exact match? */ @@ -1338,7 +1342,8 @@ __clsm_search_near(WT_CURSOR *cursor, int *exactp) if ((ret = c->next(c)) == WT_NOTFOUND) { ret = 0; continue; - } else if (ret != 0) + } + if (ret != 0) goto err; } @@ -1564,12 +1569,23 @@ __clsm_update(WT_CURSOR *cursor) WT_CURSOR_NEEDVALUE(cursor); WT_ERR(__clsm_enter(clsm, false, true)); - if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) || - (ret = __clsm_lookup(clsm, &value)) == 0) { - WT_ERR(__clsm_deleted_encode( - session, &cursor->value, &value, &buf)); - ret = __clsm_put(session, clsm, &cursor->key, &value, true); - } + if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) + WT_ERR(__clsm_lookup(clsm, &value)); + WT_ERR(__clsm_deleted_encode(session, &cursor->value, &value, &buf)); + WT_ERR(__clsm_put(session, clsm, &cursor->key, &value, true)); + + /* + * Set the cursor to reference the internal key/value of the positioned + * cursor. + */ + F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET); + WT_ITEM_SET(cursor->key, clsm->current->key); + WT_ITEM_SET(cursor->value, clsm->current->value); + WT_ASSERT(session, + F_MASK(clsm->current, WT_CURSTD_KEY_SET) == WT_CURSTD_KEY_INT); + WT_ASSERT(session, + F_MASK(clsm->current, WT_CURSTD_VALUE_SET) == WT_CURSTD_VALUE_INT); + F_SET(cursor, WT_CURSTD_KEY_INT | WT_CURSTD_VALUE_INT); err: __wt_scr_free(session, &buf); __clsm_leave(clsm); @@ -1588,18 +1604,34 @@ __clsm_remove(WT_CURSOR *cursor) WT_DECL_RET; WT_ITEM value; WT_SESSION_IMPL *session; + bool positioned; clsm = (WT_CURSOR_LSM *)cursor; + /* Check if the cursor is positioned. */ + positioned = F_ISSET(cursor, WT_CURSTD_KEY_INT); + CURSOR_REMOVE_API_CALL(cursor, session, NULL); WT_CURSOR_NEEDKEY(cursor); WT_CURSOR_NOVALUE(cursor); WT_ERR(__clsm_enter(clsm, false, true)); - if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) || - (ret = __clsm_lookup(clsm, &value)) == 0) - ret = __clsm_put( - session, clsm, &cursor->key, &__tombstone, true); + if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) + WT_ERR(__clsm_lookup(clsm, &value)); + WT_ERR(__clsm_put( + session, clsm, &cursor->key, &__tombstone, positioned)); + + /* + * If the cursor was positioned, it stays positioned with a key but no + * no value, otherwise, there's no position, key or value. This isn't + * just cosmetic, without a reset, iteration on this cursor won't start + * at the beginning/end of the table. + */ + F_CLR(cursor, WT_CURSTD_KEY_SET | WT_CURSTD_VALUE_SET); + if (positioned) + F_SET(cursor, WT_CURSTD_KEY_INT); + else + WT_TRET(cursor->reset(cursor)); err: __clsm_leave(clsm); CURSOR_UPDATE_API_END(session, ret); @@ -1692,8 +1724,8 @@ __wt_clsm_open(WT_SESSION_IMPL *session, bulk = cval.val != 0; /* Get the LSM tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, uri, bulk, &lsm_tree)); + ret = __wt_lsm_tree_get(session, uri, bulk, &lsm_tree); + /* * Check whether the exclusive open for a bulk load succeeded, and * if it did ensure that it's safe to bulk load into the tree. diff --git a/src/lsm/lsm_manager.c b/src/lsm/lsm_manager.c index cbd83a5cd30..e33e119aa41 100644 --- a/src/lsm/lsm_manager.c +++ b/src/lsm/lsm_manager.c @@ -89,7 +89,6 @@ __lsm_general_worker_start(WT_SESSION_IMPL *session) if (manager->lsm_workers % 2 == 0) FLD_SET(worker_args->type, WT_LSM_WORK_MERGE); } - F_SET(worker_args, WT_LSM_WORKER_RUN); WT_RET(__wt_lsm_worker_start(session, worker_args)); } @@ -129,17 +128,13 @@ __lsm_stop_workers(WT_SESSION_IMPL *session) manager->lsm_workers--) { worker_args = &manager->lsm_worker_cookies[manager->lsm_workers - 1]; - /* - * Clear this worker's flag so it stops. - */ - F_CLR(worker_args, WT_LSM_WORKER_RUN); - WT_ASSERT(session, worker_args->tid != 0); - WT_RET(__wt_thread_join(session, worker_args->tid)); - worker_args->tid = 0; + WT_ASSERT(session, worker_args->tid_set); + + WT_RET(__wt_lsm_worker_stop(session, worker_args)); worker_args->type = 0; - worker_args->flags = 0; + /* - * We do not clear the session because they are allocated + * We do not clear the other fields because they are allocated * statically when the connection was opened. */ } @@ -237,12 +232,12 @@ __wt_lsm_manager_start(WT_SESSION_IMPL *session) manager->lsm_worker_cookies[i].session = worker_session; } + F_SET(conn, WT_CONN_SERVER_LSM); + /* Start the LSM manager thread. */ WT_ERR(__wt_thread_create(session, &manager->lsm_worker_cookies[0].tid, __lsm_worker_manager, &manager->lsm_worker_cookies[0])); - F_SET(conn, WT_CONN_SERVER_LSM); - if (0) { err: for (i = 0; (worker_session = @@ -289,13 +284,18 @@ __wt_lsm_manager_destroy(WT_SESSION_IMPL *session) manager = &conn->lsm_manager; removed = 0; + /* + * Clear the LSM server flag and flush to ensure running threads see + * the state change. + */ + F_CLR(conn, WT_CONN_SERVER_LSM); + WT_FULL_BARRIER(); + WT_ASSERT(session, !F_ISSET(conn, WT_CONN_READONLY) || manager->lsm_workers == 0); if (manager->lsm_workers > 0) { - /* - * Stop the main LSM manager thread first. - */ - while (F_ISSET(conn, WT_CONN_SERVER_LSM)) + /* Wait for the main LSM manager thread to finish. */ + while (!F_ISSET(manager, WT_LSM_MANAGER_SHUTDOWN)) __wt_yield(); /* Clean up open LSM handles. */ @@ -303,7 +303,6 @@ __wt_lsm_manager_destroy(WT_SESSION_IMPL *session) WT_TRET(__wt_thread_join( session, manager->lsm_worker_cookies[0].tid)); - manager->lsm_worker_cookies[0].tid = 0; /* Release memory from any operations left on the queue. */ while ((current = TAILQ_FIRST(&manager->switchqh)) != NULL) { @@ -342,7 +341,7 @@ __wt_lsm_manager_destroy(WT_SESSION_IMPL *session) /* * __lsm_manager_worker_shutdown -- - * Shutdown the LSM manager and worker threads. + * Shutdown the LSM worker threads. */ static int __lsm_manager_worker_shutdown(WT_SESSION_IMPL *session) @@ -354,14 +353,13 @@ __lsm_manager_worker_shutdown(WT_SESSION_IMPL *session) manager = &S2C(session)->lsm_manager; /* - * Wait for the rest of the LSM workers to shutdown. Stop at index + * Wait for the rest of the LSM workers to shutdown. Start at index * one - since we (the manager) are at index 0. */ for (i = 1; i < manager->lsm_workers; i++) { - WT_ASSERT(session, manager->lsm_worker_cookies[i].tid != 0); - __wt_cond_signal(session, manager->work_cond); - WT_TRET(__wt_thread_join( - session, manager->lsm_worker_cookies[i].tid)); + WT_ASSERT(session, manager->lsm_worker_cookies[i].tid_set); + WT_TRET(__wt_lsm_worker_stop( + session, &manager->lsm_worker_cookies[i])); } return (ret); } @@ -383,12 +381,12 @@ __lsm_manager_run_server(WT_SESSION_IMPL *session) conn = S2C(session); dhandle_locked = false; - while (F_ISSET(conn, WT_CONN_SERVER_RUN)) { + while (F_ISSET(conn, WT_CONN_SERVER_LSM)) { __wt_sleep(0, 10000); if (TAILQ_EMPTY(&conn->lsmqh)) continue; - __wt_spin_lock(session, &conn->dhandle_lock); - F_SET(session, WT_SESSION_LOCKED_HANDLE_LIST); + __wt_readlock(session, &conn->dhandle_lock); + F_SET(session, WT_SESSION_LOCKED_HANDLE_LIST_READ); dhandle_locked = true; TAILQ_FOREACH(lsm_tree, &S2C(session)->lsmqh, q) { if (!lsm_tree->active) @@ -448,14 +446,14 @@ __lsm_manager_run_server(WT_SESSION_IMPL *session) session, WT_LSM_WORK_MERGE, 0, lsm_tree)); } } - __wt_spin_unlock(session, &conn->dhandle_lock); - F_CLR(session, WT_SESSION_LOCKED_HANDLE_LIST); + __wt_readunlock(session, &conn->dhandle_lock); + F_CLR(session, WT_SESSION_LOCKED_HANDLE_LIST_READ); dhandle_locked = false; } err: if (dhandle_locked) { - __wt_spin_unlock(session, &conn->dhandle_lock); - F_CLR(session, WT_SESSION_LOCKED_HANDLE_LIST); + __wt_readunlock(session, &conn->dhandle_lock); + F_CLR(session, WT_SESSION_LOCKED_HANDLE_LIST_READ); } return (ret); } @@ -469,11 +467,13 @@ static WT_THREAD_RET __lsm_worker_manager(void *arg) { WT_DECL_RET; + WT_LSM_MANAGER *manager; WT_LSM_WORKER_ARGS *cookie; WT_SESSION_IMPL *session; cookie = (WT_LSM_WORKER_ARGS *)arg; session = cookie->session; + manager = &S2C(session)->lsm_manager; WT_ERR(__lsm_general_worker_start(session)); WT_ERR(__lsm_manager_run_server(session)); @@ -482,7 +482,11 @@ __lsm_worker_manager(void *arg) if (ret != 0) { err: WT_PANIC_MSG(session, ret, "LSM worker manager thread error"); } - F_CLR(S2C(session), WT_CONN_SERVER_LSM); + + /* Connection close waits on us to shutdown, let it know we're done. */ + F_SET(manager, WT_LSM_MANAGER_SHUTDOWN); + WT_FULL_BARRIER(); + return (WT_THREAD_RET_VALUE); } diff --git a/src/lsm/lsm_merge.c b/src/lsm/lsm_merge.c index ceb5f03a2f5..8838638f388 100644 --- a/src/lsm/lsm_merge.c +++ b/src/lsm/lsm_merge.c @@ -187,7 +187,7 @@ __lsm_merge_span(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, continue; if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) || chunk->generation > 0) break; - else if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) && + if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) && F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) break; } @@ -625,7 +625,7 @@ err: if (locked) else __wt_verbose(session, WT_VERB_LSM, "Merge failed with %s", - __wt_strerror(session, ret, NULL, 0)); + __wt_strerror(session, ret, NULL, 0)); } F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION); return (ret); diff --git a/src/lsm/lsm_meta.c b/src/lsm/lsm_meta.c index 46ead6d6ac4..fc4dde82470 100644 --- a/src/lsm/lsm_meta.c +++ b/src/lsm/lsm_meta.c @@ -229,7 +229,7 @@ __lsm_meta_read_v1( cv.len -= 2; } WT_ERR(__wt_config_check(session, - WT_CONFIG_REF(session, WT_SESSION_create), cv.str, cv.len)); + WT_CONFIG_REF(session, WT_SESSION_create), cv.str, cv.len)); WT_ERR(__wt_strndup(session, cv.str, cv.len, &lsm_tree->bloom_config)); WT_ERR(__wt_config_getones( session, lsmconf, "lsm.bloom_hash_count", &cv)); diff --git a/src/lsm/lsm_stat.c b/src/lsm/lsm_stat.c index 150de968722..411655878af 100644 --- a/src/lsm/lsm_stat.c +++ b/src/lsm/lsm_stat.c @@ -29,24 +29,22 @@ __curstat_lsm_init( const char *cfg[] = { WT_CONFIG_BASE(session, WT_SESSION_open_cursor), NULL, NULL }; const char *disk_cfg[] = { - WT_CONFIG_BASE(session, WT_SESSION_open_cursor), - "checkpoint=" WT_CHECKPOINT, NULL, NULL }; + WT_CONFIG_BASE(session, WT_SESSION_open_cursor), + "checkpoint=" WT_CHECKPOINT, NULL, NULL }; locked = false; - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, uri, false, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, uri, false, &lsm_tree)); WT_ERR(__wt_scr_alloc(session, 0, &uribuf)); /* Propagate all, fast and/or clear to the cursors we open. */ if (cst->flags != 0) { - (void)snprintf(config, sizeof(config), + WT_ERR(__wt_snprintf(config, sizeof(config), "statistics=(%s%s%s%s)", F_ISSET(cst, WT_STAT_TYPE_ALL) ? "all," : "", F_ISSET(cst, WT_STAT_CLEAR) ? "clear," : "", !F_ISSET(cst, WT_STAT_TYPE_ALL) && F_ISSET(cst, WT_STAT_TYPE_FAST) ? "fast," : "", - F_ISSET(cst, WT_STAT_TYPE_SIZE) ? "size," : ""); + F_ISSET(cst, WT_STAT_TYPE_SIZE) ? "size," : "")); cfg[1] = disk_cfg[1] = config; } diff --git a/src/lsm/lsm_tree.c b/src/lsm/lsm_tree.c index 38d87dd852b..a9275976023 100644 --- a/src/lsm/lsm_tree.c +++ b/src/lsm/lsm_tree.c @@ -38,7 +38,7 @@ __lsm_tree_discard(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, bool final) /* We may be destroying an lsm_tree before it was added. */ if (F_ISSET(lsm_tree, WT_LSM_TREE_OPEN)) { WT_ASSERT(session, final || - F_ISSET(session, WT_SESSION_LOCKED_HANDLE_LIST)); + F_ISSET(session, WT_SESSION_LOCKED_HANDLE_LIST_WRITE)); TAILQ_REMOVE(&S2C(session)->lsmqh, lsm_tree, q); } @@ -321,9 +321,7 @@ __wt_lsm_tree_create(WT_SESSION_IMPL *session, metadata = NULL; /* If the tree can be opened, it already exists. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, uri, false, &lsm_tree)); - if (ret == 0) { + if ((ret = __wt_lsm_tree_get(session, uri, false, &lsm_tree)) == 0) { __wt_lsm_tree_release(session, lsm_tree); return (exclusive ? EEXIST : 0); } @@ -339,7 +337,7 @@ __wt_lsm_tree_create(WT_SESSION_IMPL *session, * error: the returned handle is NULL on error, and the metadata * tracking macros handle cleaning up on failure. */ - WT_WITH_HANDLE_LIST_LOCK(session, + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, ret = __lsm_tree_open(session, uri, true, &lsm_tree)); if (ret == 0) __wt_lsm_tree_release(session, lsm_tree); @@ -404,6 +402,9 @@ __lsm_tree_find(WT_SESSION_IMPL *session, } *treep = lsm_tree; + + WT_ASSERT(session, lsm_tree->excl_session == + (exclusive ? session : NULL)); return (0); } @@ -456,7 +457,8 @@ __lsm_tree_open(WT_SESSION_IMPL *session, conn = S2C(session); lsm_tree = NULL; - WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_HANDLE_LIST)); + WT_ASSERT(session, + F_ISSET(session, WT_SESSION_LOCKED_HANDLE_LIST_WRITE)); /* Start the LSM manager thread if it isn't running. */ if (__wt_atomic_cas32(&conn->lsm_manager.lsm_workers, 0, 1)) @@ -469,7 +471,7 @@ __lsm_tree_open(WT_SESSION_IMPL *session, /* Try to open the tree. */ WT_RET(__wt_calloc_one(session, &lsm_tree)); - WT_ERR(__wt_rwlock_alloc(session, &lsm_tree->rwlock, "lsm tree")); + __wt_rwlock_init(session, &lsm_tree->rwlock); WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri)); @@ -520,14 +522,21 @@ __wt_lsm_tree_get(WT_SESSION_IMPL *session, { WT_DECL_RET; - WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_HANDLE_LIST)); - - ret = __lsm_tree_find(session, uri, exclusive, treep); + /* + * Dropping and re-acquiring the lock is safe here, since the tree open + * call checks to see if another thread beat it to opening the tree + * before proceeding. + */ + if (exclusive) + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, + ret = __lsm_tree_find(session, uri, exclusive, treep)); + else + WT_WITH_HANDLE_LIST_READ_LOCK(session, + ret = __lsm_tree_find(session, uri, exclusive, treep)); if (ret == WT_NOTFOUND) - ret = __lsm_tree_open(session, uri, exclusive, treep); + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, + ret = __lsm_tree_open(session, uri, exclusive, treep)); - WT_ASSERT(session, ret != 0 || - (*treep)->excl_session == (exclusive ? session : NULL)); return (ret); } @@ -857,9 +866,7 @@ __wt_lsm_tree_alter( locked = false; /* Get the LSM tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, uri, false, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, uri, false, &lsm_tree)); /* Prevent any new opens. */ __wt_lsm_tree_writelock(session, lsm_tree); @@ -899,9 +906,7 @@ __wt_lsm_tree_drop( locked = false; /* Get the LSM tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, name, true, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, name, true, &lsm_tree)); WT_ASSERT(session, !lsm_tree->active); /* Prevent any new opens. */ @@ -934,7 +939,7 @@ __wt_lsm_tree_drop( WT_ASSERT(session, !lsm_tree->active); err: if (locked) __wt_lsm_tree_writeunlock(session, lsm_tree); - WT_WITH_HANDLE_LIST_LOCK(session, + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, tret = __lsm_tree_discard(session, lsm_tree, false)); WT_TRET(tret); return (ret); @@ -960,9 +965,7 @@ __wt_lsm_tree_rename(WT_SESSION_IMPL *session, locked = false; /* Get the LSM tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, olduri, true, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, olduri, true, &lsm_tree)); /* Prevent any new opens. */ __wt_lsm_tree_writelock(session, lsm_tree); @@ -1007,7 +1010,7 @@ err: if (locked) * Discard this LSM tree structure. The first operation on the renamed * tree will create a new one. */ - WT_WITH_HANDLE_LIST_LOCK(session, + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, tret = __lsm_tree_discard(session, lsm_tree, false)); WT_TRET(tret); return (ret); @@ -1032,9 +1035,7 @@ __wt_lsm_tree_truncate( locked = false; /* Get the LSM tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, name, true, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, name, true, &lsm_tree)); /* Prevent any new opens. */ __wt_lsm_tree_writelock(session, lsm_tree); @@ -1068,7 +1069,7 @@ err: if (locked) * the last good version of the metadata will be used, resulting * in a valid (not truncated) tree. */ - WT_WITH_HANDLE_LIST_LOCK(session, + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, tret = __lsm_tree_discard(session, lsm_tree, false)); WT_TRET(tret); } @@ -1082,7 +1083,7 @@ err: if (locked) void __wt_lsm_tree_readlock(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) { - __wt_readlock(session, lsm_tree->rwlock); + __wt_readlock(session, &lsm_tree->rwlock); /* * Diagnostic: avoid deadlocks with the schema lock: if we need it for @@ -1100,7 +1101,7 @@ __wt_lsm_tree_readunlock(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) { F_CLR(session, WT_SESSION_NO_EVICTION | WT_SESSION_NO_SCHEMA_LOCK); - __wt_readunlock(session, lsm_tree->rwlock); + __wt_readunlock(session, &lsm_tree->rwlock); } /* @@ -1110,7 +1111,7 @@ __wt_lsm_tree_readunlock(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) void __wt_lsm_tree_writelock(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) { - __wt_writelock(session, lsm_tree->rwlock); + __wt_writelock(session, &lsm_tree->rwlock); /* * Diagnostic: avoid deadlocks with the schema lock: if we need it for @@ -1128,7 +1129,7 @@ __wt_lsm_tree_writeunlock(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) { F_CLR(session, WT_SESSION_NO_EVICTION | WT_SESSION_NO_SCHEMA_LOCK); - __wt_writeunlock(session, lsm_tree->rwlock); + __wt_writeunlock(session, &lsm_tree->rwlock); } /* @@ -1157,9 +1158,7 @@ __wt_lsm_compact(WT_SESSION_IMPL *session, const char *name, bool *skipp) /* Tell __wt_schema_worker not to look inside the LSM tree. */ *skipp = true; - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, name, false, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, name, false, &lsm_tree)); if (!F_ISSET(S2C(session), WT_CONN_LSM_MERGE)) WT_ERR_MSG(session, EINVAL, @@ -1356,9 +1355,7 @@ __wt_lsm_tree_worker(WT_SESSION_IMPL *session, locked = false; exclusive = FLD_ISSET(open_flags, WT_DHANDLE_EXCLUSIVE); - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __wt_lsm_tree_get(session, uri, exclusive, &lsm_tree)); - WT_RET(ret); + WT_RET(__wt_lsm_tree_get(session, uri, exclusive, &lsm_tree)); /* * We mark that we're busy using the tree to coordinate diff --git a/src/lsm/lsm_work_unit.c b/src/lsm/lsm_work_unit.c index d9c185a3f58..e6a29666094 100644 --- a/src/lsm/lsm_work_unit.c +++ b/src/lsm/lsm_work_unit.c @@ -230,7 +230,7 @@ __wt_lsm_work_bloom(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) if (__wt_atomic_cas32(&chunk->bloom_busy, 0, 1)) { if (!F_ISSET(chunk, WT_LSM_CHUNK_BLOOM)) { ret = __lsm_bloom_create( - session, lsm_tree, chunk, (u_int)i); + session, lsm_tree, chunk, i); /* * Record if we were successful so that we can * later push a merge work unit. @@ -265,9 +265,9 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, { WT_DECL_RET; WT_TXN_ISOLATION saved_isolation; - bool flush_set; + bool flush_set, release_btree; - flush_set = false; + flush_set = release_btree = false; /* * If the chunk is already checkpointed, make sure it is also evicted. @@ -276,7 +276,7 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, if (F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !F_ISSET(chunk, WT_LSM_CHUNK_STABLE) && !chunk->evicted) { - WT_WITH_HANDLE_LIST_LOCK(session, + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, ret = __lsm_discard_handle(session, chunk->uri, NULL)); if (ret == 0) chunk->evicted = 1; @@ -318,20 +318,18 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, * 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, WT_SYNC_WRITE_LEAVES); - session->txn.isolation = saved_isolation; - WT_TRET(__wt_session_release_btree(session)); - } - WT_ERR(ret); + WT_ERR(__wt_session_get_btree(session, chunk->uri, NULL, NULL, 0)); + release_btree = true; + + /* + * 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; + WT_ERR(__wt_cache_op(session, WT_SYNC_WRITE_LEAVES)); + session->txn.isolation = saved_isolation; __wt_verbose(session, WT_VERB_LSM, "LSM worker checkpointing %s", chunk->uri); @@ -348,12 +346,14 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, WT_ERR(__wt_meta_track_on(session)); WT_WITH_CHECKPOINT_LOCK(session, WT_WITH_SCHEMA_LOCK(session, - ret = __wt_schema_worker( - session, chunk->uri, __wt_checkpoint, NULL, NULL, 0))); + ret = __wt_checkpoint(session, NULL))); WT_TRET(__wt_meta_track_off(session, false, ret != 0)); if (ret != 0) WT_ERR_MSG(session, ret, "LSM checkpoint"); + release_btree = false; + WT_ERR(__wt_session_release_btree(session)); + /* Now the file is written, get the chunk size. */ WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk)); @@ -376,16 +376,6 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, 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_lsm_switch_primary(session, false); - WT_ERR(__wt_session_release_btree(session)); - /* Make sure we aren't pinning a transaction ID. */ __wt_txn_release_snapshot(session); @@ -402,6 +392,8 @@ __wt_lsm_checkpoint_chunk(WT_SESSION_IMPL *session, err: if (flush_set) WT_PUBLISH(chunk->flushing, 0); + if (release_btree) + WT_TRET(__wt_session_release_btree(session)); return (ret); } @@ -517,8 +509,8 @@ __lsm_drop_file(WT_SESSION_IMPL *session, const char *uri) * * This will fail with EBUSY if the file is still in use. */ - WT_WITH_HANDLE_LIST_LOCK(session, - ret = __lsm_discard_handle(session, uri, WT_CHECKPOINT)); + WT_WITH_HANDLE_LIST_WRITE_LOCK(session, + ret = __lsm_discard_handle(session, uri, WT_CHECKPOINT)); WT_RET(ret); /* @@ -610,7 +602,8 @@ __wt_lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) if (drop_ret == EBUSY) { ++skipped; continue; - } else if (drop_ret != ENOENT) + } + if (drop_ret != ENOENT) WT_ERR(drop_ret); flush_metadata = true; @@ -621,7 +614,8 @@ __wt_lsm_free_chunks(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree) if (drop_ret == EBUSY) { ++skipped; continue; - } else if (drop_ret != ENOENT) + } + if (drop_ret != ENOENT) WT_ERR(drop_ret); flush_metadata = true; } diff --git a/src/lsm/lsm_worker.c b/src/lsm/lsm_worker.c index b0d0758775d..1cabbd4888d 100644 --- a/src/lsm/lsm_worker.c +++ b/src/lsm/lsm_worker.c @@ -21,7 +21,23 @@ __wt_lsm_worker_start(WT_SESSION_IMPL *session, WT_LSM_WORKER_ARGS *args) { __wt_verbose(session, WT_VERB_LSM_MANAGER, "Start LSM worker %u type %#" PRIx32, args->id, args->type); - return (__wt_thread_create(session, &args->tid, __lsm_worker, args)); + + args->running = true; + WT_RET(__wt_thread_create(session, &args->tid, __lsm_worker, args)); + args->tid_set = true; + return (0); +} + +/* + * __wt_lsm_worker_stop -- + * A wrapper around the LSM worker thread stop. + */ +int +__wt_lsm_worker_stop(WT_SESSION_IMPL *session, WT_LSM_WORKER_ARGS *args) +{ + args->running = false; + args->tid_set = false; + return (__wt_thread_join(session, args->tid)); } /* @@ -84,7 +100,6 @@ err: __wt_lsm_manager_free_work_unit(session, entry); static WT_THREAD_RET __lsm_worker(void *arg) { - WT_CONNECTION_IMPL *conn; WT_DECL_RET; WT_LSM_WORK_UNIT *entry; WT_LSM_WORKER_ARGS *cookie; @@ -93,11 +108,9 @@ __lsm_worker(void *arg) cookie = (WT_LSM_WORKER_ARGS *)arg; session = cookie->session; - conn = S2C(session); entry = NULL; - while (F_ISSET(conn, WT_CONN_SERVER_RUN) && - F_ISSET(cookie, WT_LSM_WORKER_RUN)) { + while (cookie->running) { progress = false; /* @@ -154,7 +167,7 @@ __lsm_worker(void *arg) /* Don't busy wait if there was any work to do. */ if (!progress) { - __wt_cond_wait(session, cookie->work_cond, 10000); + __wt_cond_wait(session, cookie->work_cond, 10000, NULL); continue; } } |