| Bug #27370 | Potential inconsistent blob reads for ReadCommitted reads | ||
|---|---|---|---|
| Submitted: | 22 Mar 2007 12:15 | Modified: | 30 Apr 2007 9:40 |
| Reporter: | Kristian Nielsen | Email Updates: | |
| Status: | Closed | Impact on me: | |
| Category: | MySQL Cluster: NDB API | Severity: | S2 (Serious) |
| Version: | 5.0-bk, 5.1-bk | OS: | Any (Any) |
| Assigned to: | Kristian Nielsen | CPU Architecture: | Any |
[26 Mar 2007 10:18]
Kristian Nielsen
I think this bug only affects NDB API programs, as the mysqld ha_ndbcluster.cc
seems to set the lock mode to LM_Read itself for blob access:
int ha_ndbcluster::get_ndb_lock_type(enum thr_lock_type type)
{
if (type >= TL_WRITE_ALLOW_WRITE)
return(NdbOperation::LM_Exclusive);
else if (type == TL_READ_WITH_SHARED_LOCKS ||
uses_blob_value(m_retrieve_all_fields))
return(NdbOperation::LM_Read);
else
return(NdbOperation::LM_CommittedRead);
}
So only NDB API programs that read blobs with non-default LM_CommittedRead would be affected.
[28 Mar 2007 9:33]
Bugs System
A patch for this bug has been committed. After review, it may be pushed to the relevant source trees for release in the next version. You can access the patch from: http://lists.mysql.com/commits/23125 ChangeSet@1.2417, 2007-03-28 11:29:13+02:00, knielsen@ymer.(none) +7 -0 BUG#27370: Potential inconsistent blob reads for ReadCommitted reads. The old blob implementation had code that attempted to upgrade the lock mode for LM_CommittedRead operations, but it did not work properly as it did not recompute the operation flags properly. As a consequence, reading a blob with LM_CommittedRead could return inconsistent data, with different part of the read data being from different commits done by other transactions. The fix is to correctly recompute all necessary flags when upgrading lock mode.
[24 Apr 2007 6:19]
Bugs System
A patch for this bug has been committed. After review, it may be pushed to the relevant source trees for release in the next version. You can access the patch from: http://lists.mysql.com/commits/25224 ChangeSet@1.2417, 2007-04-24 08:15:31+02:00, knielsen@ymer.(none) +8 -0 BUG#27370: Potential inconsistent blob reads for ReadCommitted reads. The old blob implementation had code that attempted to upgrade the lock mode for LM_CommittedRead operations, but it did not work properly as it did not recompute the operation flags. As a consequence, reading a blob with LM_CommittedRead could return inconsistent data, with different part of the read data being from different commits done by other transactions. The fix is to correctly recompute all necessary flags when upgrading lock mode.
[25 Apr 2007 7:30]
Bugs System
A patch for this bug has been committed. After review, it may be pushed to the relevant source trees for release in the next version. You can access the patch from: http://lists.mysql.com/commits/25370 ChangeSet@1.2401, 2007-04-25 09:30:15+02:00, knielsen@ymer.(none) +8 -0 BUG#27370: Potential inconsistent blob reads for ReadCommitted reads. The old blob implementation had code that attempted to upgrade the lock mode for LM_CommittedRead operations, but it did not work properly as it did not recompute the operation flags. As a consequence, reading a blob with LM_CommittedRead could return inconsistent data, with different part of the read data being from different commits done by other transactions. The fix is to correctly recompute all necessary flags when upgrading lock mode.
[25 Apr 2007 19:11]
Kristian Nielsen
Pushed to mysql-5.0-ndb and mysql-5.1-new-ndb.
[28 Apr 2007 19:35]
Bugs System
Pushed into 5.1.18-beta
[28 Apr 2007 19:36]
Bugs System
Pushed into 5.0.42
[30 Apr 2007 9:40]
Jon Stephens
Thank you for your bug report. This issue has been committed to our source repository of that product and will be incorporated into the next release.
If necessary, you can access the source repository and build the latest available version, including the bug fix. More information about accessing the source trees is available at
http://dev.mysql.com/doc/en/installing-source.html
Documented bugfix in 5.0.42 and 5.1.18 changelogs.
Description: For blob reads on operations with lock mode LM_CommittedRead, the NdbBlob code is supposed to upgrade the lock mode to LM_Read. This to avoid inconstistent blob reads, where half the read see data committed by another transaction, and the other half does not. However, the code does not work. It changes the lock mode, but at that point, the NdbOperation::readTuple() method has already computed these fields from the old lock mode: NdbTransaction::theSimpleState NdbOperation::theSimpleIndicator NdbOperation::theDirtyIndicator Since these are not recomputed, the old lock mode is the one that is send in the signal to the kernel. The situation is similar for NdbScanOperation::readTuples() and NdbIndexScanOperation::readTuples(). How to repeat: Source code inspection, or check a signal trace from testBlobs. Suggested fix: Will commit a patch.