Full_Name: Lorenz Bauer Version: git c367c1f69685 OS: OS X URL: ftp://ftp.openldap.org/incoming/ Submission from: (NULL) (2a06:98c0:1000:1200:156f:df88:ee9a:7775)
The function mdb_env_copyfd2 with MDB_CP_COMPACT enabled can currently deadlock, since a call to mdb_env_cthr_toggle is not checked. The testcase at https://gist.github.com/lmb/17a528cdadde73fb231a2a1ed84714f7 reproduces this issue as follows: - Use pthread_sigmask to ignore SIGPIPE - Create new pipe - Spawn thread with mdb_env_copyfd2 using MDB_CP_COMPACT - Close read side of the pipe in main thread - Join thread and check rc code On LMDB master the testase e hangs instead of returning EPIPE. The following patch fixes the deadlock, but since I'm not very familiar with the codebase it's likely incorrect. diff --git a/libraries/liblmdb/mdb.c b/libraries/liblmdb/mdb.c index 79a958b..97ff7c7 100644 --- a/libraries/liblmdb/mdb.c +++ b/libraries/liblmdb/mdb.c @@ -9988,11 +9988,15 @@ mdb_env_copyfd1(MDB_env *env, HANDLE fd) rc = mdb_env_cwalk(&my, &txn->mt_dbs[MAIN_DBI].md_root, 0); if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle]) rc = mdb_env_cthr_toggle(&my, 1); - mdb_env_cthr_toggle(&my, -1); + rc = mdb_env_cthr_toggle(&my, -1); + if (rc) + goto done; pthread_mutex_lock(&my.mc_mutex); while(my.mc_new) pthread_cond_wait(&my.mc_cond, &my.mc_mutex); pthread_mutex_uocock(&my.mc_mutex); + +done: THREAD_FINISH(thr); mdb_txn_abort(txn);
