Additional SQL Server diagnostics added to detect unreported I/O problems

If operating system, driver, or hardware problems cause lost write conditions or stale read conditions, you may see data integrity-related error messages such as Errors 605, 823, 3448. You may receive error messages that are similar to the following examples:

Service pack information

To resolve this problem, obtain the latest service pack for Microsoft SQL Server 2000. For additional information, click the following article number to view the article in the Microsoft Knowledge Base:

Build information

Microsoft has released a build of SQL Server that introduces extended tracking capabilities. These capabilities are designed to help you troubleshoot the error messages described in the "Symptoms" section. If you do not suspect an operating system or a hardware releated problem, you may not have to apply this build.

The English version of this build has the file attributes (or later) that are listed in the following table. The dates and times for these files are listed in coordinated universal time (UTC). When you view the file information, it is converted to local time. To find the difference between UTC and local time, use the Time Zone tab in the Date and Time tool in Control Panel. Note Because of file dependencies, the most recent hotfix or feature that contains these files may also contain additional files.

Microsoft has confirmed that under rare and heavy I/O loads, some hardware platforms can return a stale read. If the extended diagnostics indicate a possible stale read/lost write condition, contact your hardware vendor for immediate follow up and test with the SQLIOStress utility.

Microsoft has confirmed that this is a problem in the Microsoft products that are listed in the "Applies to" section.

This problem was first corrected in Microsoft SQL Server 2000 Service Pack 4.

If you receive any of the error messages that are mentioned in the "Symptoms" section and they cannot be explained by an event such as a physical drive failure, then review any known problems with SQL Server, the operating system, the drivers, and the hardware. Additional SQL Server diagnostics have been added in this build to help detect I/O related external problems. The additional diagnostics try to provide information about the two following conditions:

• Lost Write: A successful call to the WriteFile API, but the operating system, a driver, or the caching controller does not correctly flush the data to the physical media even though SQL Server is informed that the write was successful.
• Stale Read: A successful call to the ReadFile API, but the operating system, a driver, or the caching controller incorrectly returns an older version of the data.

For example, Microsoft has confirmed scenarios where a WriteFile API call returns as successful, but an immediate, successful read of the same data block returns older data, including data that is likely stored in a hardware read cache. Sometimes, this problem occurs because of a read cache problem. In other cases, the write data is never actually written to the physical disk.

To enable additional diagnostics for these types of problems, SQL Server has added trace flag 818. You can specify trace flag 818 as a startup parameter, -T818, for the computer that is running SQL Server, or you can run the following statement:
Trace flag 818 enables an in-memory ring buffer that is used for tracking the last 2,048 successful write operations that are performed by the computer running SQL Server, not including sort and workfile I/Os. When errors such as Error 605, 823, or 3448 occur, the incoming buffer's log sequence number (LSN) value is compared to the recent write list. If the LSN that is retrieved during the read operation is older than the one specified during the write operation, a new error message is logged in the SQL Server error log. Most SQL Server write operations occur as checkpoints or as lazy writes. A lazy write is a background task that uses asynchronous I/O. The implementation of the ring buffer is lightweight, thereby making the performance affect on the system negligible.

The following message indicates that SQL Server did not receive an error from the WriteFile API call or the ReadFile API call. However, when the LSN was reviewed, the value was not correct: At this point, either the read cache contains an older version of the page, or the data was not correctly written to the physical disk. In either case (a Lost Write or a Stale Read), SQL Server reports an external problem with the operating system, the driver, or the hardware layers.

If Error 3448 occurs when you try to rollback a transaction that has Error 605 or Error 823, the computer running SQL Server automatically closes the database and tries to open and recover the database. The first page that experiences Error 605 or Error 823 is considered a bad page, and the page id is kept by the computer running SQL Server. During recovery (before the redo phase) when the bad page id is read, the primary details about the page header are logged in the SQL Server error log. This action is important because it helps to distinguish between Lost Write and Stale Read scenarios.

You may see the following two common behaviors in Stale Read scenarios:

• If the database files are closed and then opened, the correct and most recently written data is returned during recovery.
• When you issue a checkpoint and run the DBCC DROPCLEANBUFFERS statement (to remove all database pages from the memory), and then run the DBCC CHECKDB statement on the database, the most recently written data is returned.

The behaviors mentioned in the previous paragraph indicate a read caching problem and they are frequently solved by disabling the read cache. The actions that are outlined in the previous paragraph typically force a cache invalidation and the successful reads that occur show that the physical media is correctly updated. The Lost Write behavior occurs when the page that is read back is still the older version of the data, even after a forced flush of the caching mechanisms.

Sometimes, the problem may not be specific to a hardware cache. It may be a problem with a filter driver. In such cases, review your software, including backup utilities and antivirus software, and then see if there are problems with the filter driver.

Microsoft has also noted conditions that do not meet the criteria for Error 605 or Error 823 but are caused by the same Stale Read or Lost Write activity. In some instances, a page appears to be updated two times but with the same LSN value. This behavior may occur if the Object ID and the Page ID are correct (page already allocated to the object), and a change is made to the page and flushed to the disk. The next page retrieval returns an older image, and then a second change is made. The SQL Server transaction log shows that the page was updated two times with the same LSN value. This action becomes a problem when you try to restore a transaction log sequence or with data consistency problems, such as foreign key failures or missing data entries. The following error message illustrates one example of this condition:
Some scenarios are outlined in more detail in the following lists:

• LSN Sequence		Action
  1			Checkpoint
  2			Begin Transaction
  3			Table created or truncated
  4			Inserts (Pages allocated)
  5			Newly allocated page written to disk by Lazy Writer
  6			Select from table – Scans IAM chain, newly allocated page read back from disk (LRU | HASHED = 0x9 in getpage message), encounters Error 605 - Invalid Object ID
  7			Rollback of transaction initiated
  

• LSN Sequence		Action
  1			Checkpoint
  2			Begin Transaction
  3			Page Modification
  4			Page written to disk by Lazy Writer
  5			Page read in for another modification (stale image returned)
  6			Page Modified for a second time but because of stale image does not see first modification 
  7			Rollback – Fails – Transaction Log shows two different log records with the same PREV LSN for the page
  

SQL Server 'sort' operators perform I/O activities, primarily to and from the tempdb database. These I/O operations are similar to the buffer I/O operations; however, they have already been designed to use read retry logic to try to resolve similar issues. The additional diagnostics explained in this article do not apply to these I/O operations.

Microsoft has noted that the root cause for the following sort read failures is generally a Stale Read or a Lost Write:Customers who have experienced these sort errors have frequently resolved the problems by moving tempdb to a non-caching local drive, or by disabling the read caching mechanisms.

Because a stale read or a lost write results in data storage that is not expected, a wide variety of behaviors may occur. It may appear as missing data, but some of the more common effects of missing data appear as index corruptions, such as Error 644 or Error 625:Some customers have reported missing rows after they perform row count activities. This problem occurs because of a lost write. Perhaps the page was supposed to be linked to the clustered index page chain. If the write was physically lost, the data is also lost.

Important If you experience any of the behaviors, or if you are suspicious of similar problems together with disabling caching mechanisms, Microsoft strongly recommends that you obtain the latest update for SQL Server and the latest SQL Server I/O Stress Simulator. Microsoft also strongly encourages that you perform a strict review of your operating system and its associated configurations.

For more information, click the following article number to view the article in the Microsoft Knowledge Base:

SQL Server requires systems to support ‘ guaranteed delivery to stable media ’ as outlined under the Microsoft SQL Server Always-On Storage Solution Review program. Fo For more information about the input and output requirements for the SQL Server database engine, click the following article number to view the article in the Microsoft Knowledge Base: