From be27946d84dbf9155d9505d00abe91089a9b8125 Mon Sep 17 00:00:00 2001 From: Andrew Tridgell Date: Mon, 26 Sep 2005 01:12:12 +0000 Subject: r10492: work around a bug in solaris which cases lock upgrades to fail with EDEADLK even when progress can be made. This is not a good solution, but I can't find anything better. (This used to be commit 980dd17f7d0a622cd772afc9ba15e50007ad9c6e) --- source4/lib/tdb/common/lock.c | 26 ++++++++++++++++++++++++++ source4/lib/tdb/common/tdb_private.h | 1 + source4/lib/tdb/common/transaction.c | 2 +- 3 files changed, 28 insertions(+), 1 deletion(-) (limited to 'source4/lib') diff --git a/source4/lib/tdb/common/lock.c b/source4/lib/tdb/common/lock.c index 703cfe9dc5..04135a2274 100644 --- a/source4/lib/tdb/common/lock.c +++ b/source4/lib/tdb/common/lock.c @@ -81,6 +81,32 @@ int tdb_brlock_len(struct tdb_context *tdb, tdb_off_t offset, } +/* + upgrade a read lock to a write lock. This needs to be handled in a + special way as some OSes (such as solaris) have too conservative + deadlock detection and claim a deadlock when progress can be + made. For those OSes we may loop for a while. +*/ +int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len) +{ + int count = 1000; + while (count--) { + struct timeval tv; + if (tdb_brlock_len(tdb, offset, F_WRLCK, F_SETLKW, 1, len) == 0) { + return 0; + } + if (errno != EDEADLK) { + break; + } + /* sleep for as short a time as we can - more portable than usleep() */ + tv.tv_sec = 0; + tv.tv_usec = 1; + select(0, NULL, NULL, NULL, &tv); + } + return -1; +} + + /* a byte range locking function - return 0 on success this functions locks/unlocks 1 byte at the specified offset. diff --git a/source4/lib/tdb/common/tdb_private.h b/source4/lib/tdb/common/tdb_private.h index d7471537d8..ce105a2a56 100644 --- a/source4/lib/tdb/common/tdb_private.h +++ b/source4/lib/tdb/common/tdb_private.h @@ -216,6 +216,7 @@ void tdb_mmap(struct tdb_context *tdb); int tdb_lock(struct tdb_context *tdb, int list, int ltype); int tdb_unlock(struct tdb_context *tdb, int list, int ltype); int tdb_brlock(struct tdb_context *tdb, tdb_off_t offset, int rw_type, int lck_type, int probe); +int tdb_brlock_upgrade(struct tdb_context *tdb, tdb_off_t offset, size_t len); int tdb_brlock_len(struct tdb_context *tdb, tdb_off_t offset, int rw_type, int lck_type, int probe, size_t len); int tdb_write_lock_record(struct tdb_context *tdb, tdb_off_t off); diff --git a/source4/lib/tdb/common/transaction.c b/source4/lib/tdb/common/transaction.c index cf0f378dce..9d37046116 100644 --- a/source4/lib/tdb/common/transaction.c +++ b/source4/lib/tdb/common/transaction.c @@ -772,7 +772,7 @@ int tdb_transaction_commit(struct tdb_context *tdb) } /* upgrade the main transaction lock region to a write lock */ - if (tdb_brlock_len(tdb, FREELIST_TOP, F_WRLCK, F_SETLKW, 0, 0) == -1) { + if (tdb_brlock_upgrade(tdb, FREELIST_TOP, 0) == -1) { TDB_LOG((tdb, 0, "tdb_transaction_start: failed to upgrade hash locks\n")); tdb->ecode = TDB_ERR_LOCK; tdb_transaction_cancel(tdb); -- cgit