| Bug #48197 | Concurrent rw_lock_free may cause assertion failure | ||
|---|---|---|---|
| Submitted: | 21 Oct 9:44 | Modified: | 23 Oct 13:27 |
| Reporter: | Yasufumi Kinoshita | ||
| Status: | Verified | ||
| Category: | Server: InnoDB | Severity: | S3 (Non-critical) |
| Version: | 5.0, 5.1, Plugin-1.0.4 | OS: | Any |
| Assigned to: | Marko Mäkelä | Target Version: | |
| Tags: | Contribution | ||
| Triage: | Triaged: D4 (Minor) | ||
[22 Oct 6:58]
Yasufumi Kinoshita
It is problem for all InnoDB.
[23 Oct 13:27]
Miguel Solorzano
Thank you for the bug report.
Description: "lock->magic_n = 0" should not be done before the lock removed from the lock list. The other rw_lock_free() may cause the assertion error about the magic_n. How to repeat: It may be rare case. Suggested fix: move "lock->magic_n = 0;" to after the removing from rw_lock_list like mutex_free() does. --- innodb_plugin-1.0.4/sync/sync0rw.c 2009-05-26 22:28:49.000000000 +0900 +++ innodb_plugin-1.0.4/sync/sync0rw.c 2009-10-20 15:55:22.000000000 +0900 @@ -304,8 +304,6 @@ ut_ad(rw_lock_validate(lock)); ut_a(lock->lock_word == X_LOCK_DECR); - lock->magic_n = 0; - #ifndef INNODB_RW_LOCKS_USE_ATOMICS mutex_free(rw_lock_get_mutex(lock)); #endif /* INNODB_RW_LOCKS_USE_ATOMICS */ @@ -325,6 +323,8 @@ UT_LIST_REMOVE(list, rw_lock_list, lock); mutex_exit(&rw_lock_list_mutex); + + lock->magic_n = 0; } #ifdef UNIV_DEBUG