Microsoft KB Archive/169960

= INF: Analyzing and Avoiding Deadlocks in SQL Server =

Article ID: 169960

Article Last Modified on 10/16/2003

-

APPLIES TO


 * Microsoft SQL Server 6.5 Standard Edition

-



This article was previously published under Q169960



SUMMARY
Microsoft SQL Server maintains transactional integrity and database consistency by using locks. SQL Server version 6.5 optionally uses row- level locking for insert operations and uses page-level locking for other operations. As with any relational database system, locking may lead to deadlocks between users.

For example, suppose User1 (or Connection1) has a lock on data item "A" and wants a lock on data item "B." User2 has a lock on data item "B" and now wants a lock on data item "A." In this SQL Server scenario, either User1 or User2 will be a deadlock victim, and the other user will be granted the requested lock.

In SQL Server, the application developer can decide which connection will be the candidate for deadlock victim by using SET DEADLOCK_PRIORITY. If the developer does not designate a priority for deadlocks, SQL Server selects the deadlock victim by choosing the process that completes the circular chain of locks.

Database application systems may behave differently when ported from one relational database to another, based on the implementation of the relational database system. One of the areas to look for behavioral changes is locking. This article explains how to analyze the deadlocks in SQL Server and the techniques you can use to avoid them.



MORE INFORMATION
This article emphasizes using the output of trace flag T1204 to analyze deadlocks. When trace flag T1204 is set, SQL Server prints information about the deadlock when it occurs. To use this trace flag, use the following command at a command prompt to start SQL Server:

sqlservr -c -T1204

The trace results are sent to the console window, unless you set trace flag T3605, which sends the trace output to the error log.

Deadlocks can occur when two connections update tables in opposite order. For example, one connection inserts into table "example1" first and then into "example2," while another connection inserts into table "example2" first and then into "example1" within a transaction. An example scenario is useful to illustrate how to avoid deadlocks.

The following are the SQL statements used to create the table used for this example:

create table example1 (column1 int, column2 char(20), column3 char(50)) go  create table example2 (column1 int, column2 char(20), column3 char(50)) go  declare @lvar int select @lvar = 0 while @lvar < 500 begin insert into example1 values (@lvar, 'AAA', 'CCC') insert into example2 values (@lvar, 'AAA', 'CCC') select @lvar = @lvar + 1 end go  create unique clustered index ex1ind1 on example1 (column1, column2) with fill factor = 90, PAD_INDEX go  create unique clustered index ex2ind1 on example2 (column1, column2) with fill factor = 90, PAD_INDEX go

Example 1: Table Insertions in Opposite Order
In this example, two tables were inserted in opposite order and a deadlock occurred. Deadlocks can also occur when two or more connections perform updates or deletes on tables in opposite order.

  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC') Connection2 > INSERT INTO example2 VALUES (200, 'AAAB', 'CCC') Connection2 > INSERT INTO example1 VALUES (200, 'AAAB', 'CCC')

At this point, Connection1 may block Connection2 because the row Connection2 is inserting may be on the same page where Connection1 has already inserted a row and is holding a lock.

  Connection1 > INSERT INTO example2 VALUES (100, 'AAAA', 'CCC')

At this point, Connection2 may block Connection1, because the row Connection1 is inserting may be on the same page where Connection2 has already inserted a row and is holding a lock. This causes a deadlock.

The following is the output for trace flag 1204 when the deadlock occurred:

97/04/20 11:51:57.88 spid13  *** DEADLOCK DETECTED with spid 14 *** spid 13 requesting EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 14, dbid 6, page 0x188, table example2, indid 0x1 pcurcmd INSERT(0xc3), input buffer: INSERT INTO example2 VALUES (100,    'AAAA', 'CCC') spid 14 waiting for EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 13, dbid 6, page 0x180, table example1, indid 0x1 pcurcmd INSERT(0xc3), input buffer: INSERT INTO example1 VALUES (200,  'AAAB', 'CCC') VICTIM: spid 13, pstat 0x0000, cputime 30

Each line of the deadlock trace can tell users more about a deadlock. Connection1 is spid 13 and Connection2 is spid 14 (you can determine the spid associated with a connection by using the sp_who system stored procedure).

  >> 97/04/20 11:51:57.88 spid13   *** DEADLOCK DETECTED with spid 14 *** The deadlock was detected between spid 13 and spid 14. >> spid 13 requesting EX_PAGE (waittype 0x8005), blocked by: >>  EX_PAGE: spid 14, dbid 6, page 0x188, table example2, indid 0x1 >>  pcurcmd INSERT(0xc3), input buffer: INSERT INTO example2 VALUES (100, 'AAAA', 'CCC')

Spid 13 was requesting EX_PAGE lock and was blocked by spid 14, which already has EX_PAGE lock for page 0x188 on table example2 in dbid 6. The lock is held on the page belonging to clustered index.

     Indid Value         Description -        0                Data page if there is no clustered index, or the leaf page of a clustered index if there is one 1               Non-leaf page of the clustered index page 255               Text/image page Any other value      Non-clustered secondary index

The current command executed by spid 13 is an INSERT and the trace gives part of the input buffer.

  >> spid 14 waiting for EX_PAGE (waittype 0x8005), blocked by: >>  EX_PAGE: spid 13, dbid 6, page 0x180, table example1, indid 0x1 >>  pcurcmd INSERT(0xc3), input buffer: INSERT INTO example1 VALUES (200, 'AAAB', 'CCC')

Spid 14 is waiting for EX_PAGE lock and is being blocked by spid 13, which already holds EX_PAGE lock on the same page.

  >> VICTIM: spid 13, pstat 0x0000, cputime 30 SQL Server has chosen spid 13 as the deadlock victim.

The following is an explanation of what the various locks mean in the trace:

SH_INT and EX_INT

Intent locks that are taken on a higher-level item (for example, a table) before lower-level locks (for example, a page) can be taken, because the lock manager is unaware of the relationship between different types of items (in this case, pages and tables). If an EX_INT lock was not taken on the table before taking EX_PAG locks on the pages, another user could take an EX_TAB lock on the same table and the lock manager would not know that a conflict existed. Currently, SQL Server has intent locks only on tables. There are two kinds of intent locks: shared (SH_INT) and exclusive (EX_INT) locks.

EX_PAGE

This is an exclusive page lock that is taken when a page is updated due to a DELETE, UPDATE, or INSERT statement with insert row-level locking (IRL) disabled.

UP_PAGE

This is an update page lock that is taken in place of a shared-page lock when a page is scanned and the optimizer knows that the page will be updated (or the UPDLOCK hint is used).

PR_EXT, NX_EXT, UPD_EXT, and EX_EXT

These locks are taken when allocating or deallocating disk space. UPD_EXT is taken when allocating or deallocating a page from an existing extent and the others are used when allocating or deallocating entire extents.

IX_PAGE and LN_PAGE

These are IRL locks. IX_PAGE is an intent-to-do-row-locking lock on a page. LN_PAGE is taken when a page on which IRL is being done needs to be split.

RLOCK and XRLOCK

These short-term locks are taken when traversing an index b-tree. There are two types of this kind of lock: shared (RLOCK) and exclusive (XRLOCK). Shared locks are taken during scan, while exclusive locks are taken on index pages during an update.

EX_TAB

This is an exclusive table lock that occurs when the SQL Server optimizer determines that a table scan is the most efficient way to solve an update query (for example, when there are no indexes on a table). EX_TAB locks also appear when you lock the table with TABLOCKX hint or when SQL Server escalates the page locks on a table to a table lock.

SH_TAB

This is a shared table lock that is used when the optimizer assumes that most of the table will be scanned (or page locking escalates) or the TABLOCK hint is used.

The previous deadlock example can be avoided if the two connections update tables in the following sequence:

  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC') Connection2 > INSERT INTO example1 VALUES (200, 'AAAB', 'CCC') Connection2 > INSERT INTO example2 VALUES (200, 'AAAB', 'CCC') Connection1 > INSERT INTO example2 VALUES (100, 'AAAA', 'CCC')

Example 2: Insertions to Different Parts of the Same Table
This deadlock can also occur when two connections insert into different parts of the same table in opposite order when rows share pages. For example:

  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC') Connection2 > INSERT INTO example1 VALUES (400, 'AAAB', 'CCC') Connection1 > INSERT INTO example1 VALUES (400, 'AAAA', 'CCC')

In this example table, there is a clustered index on the first column of the example1 table. Rows with the same values for the first column will tend to fall on the same page. In the example, the second row inserted by Connection1 will probably fall on the same page as the first row inserted by Connection2, because they both have a clustered index value of 400. This causes Connection2 to block Connection1.

  Connection2 > INSERT INTO example1 VALUES (100, 'AAAB', 'CCC')

Now Connection2 may also be blocked by Connection1, leading to a deadlock. The following is the deadlock trace:

  97/04/20 12:56:01.40 spid16   *** DEADLOCK DETECTED with spid 15 *** spid 16 requesting EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 15, dbid 6, page 0x2c5, table example1, indid 0 pcurcmd INSERT(0xc3), input buffer: INSERT INTO example1 VALUES (100,  'AAAB', 'CCC') spid 15 waiting for EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 16, dbid 6, page 0x8bd, table example1, indid 0 pcurcmd INSERT(0xc3), input buffer: INSERT INTO example1 VALUES (400,  'AAAA', 'CCC') VICTIM: spid 16, pstat 0x0000, cputime 130

The spid 16 request for EX_PAGE lock for page 0x2c5 is blocked by spid 15, which already holds EX_PAGE lock for page 0x2c5 after it did the first insert. And spid 15 also got blocked by spid 16 on waiting for a EX_PAGE lock for page 0x8db leading to deadlock.

This deadlock can be avoided by using the following command to enable IRL for table example1:

  sp_tableoption 'example1', 'insert row lock', true

Example 3: Insertions Using IRL
IRL allows two or more users to share a page when they do only insert operations, which often results in better throughput. However, enabling IRL will not always reduce deadlocks. In some cases, IRL may introduce deadlocks.

  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC') Connection2 > INSERT INTO example1 VALUES (105, 'AAAB', 'CCC')

With IRL enabled, both connections will hold an IX_PAGE lock on the page containing the two new rows. If IRL was disabled, Connection1 would have acquired an EX_PAGE lock, and Connection2 would have been blocked immediately.

<pre class="fixed_text">  Connection2 > UPDATE example1 SET column3 = 'CCCB' where column1 = 105 and column2 = 'AAAB'

At this point, Connection2 needs an exclusive page lock to do an UPDATE statement, which is incompatible with Connection1's IX_PAGE lock. Therefore, Connection2 will wait.

<pre class="fixed_text">  Connection1 > UPDATE example1 SET column3 = 'CCCA' where column1 = 100 and column2 = 'AAAA'

Now Connection1 may be blocked by Connection2, leading to a deadlock. The following is the deadlock trace:

<pre class="fixed_text">  97/04/20 15:13:50.07 spid17   *** DEADLOCK DETECTED with spid 18 *** spid 17 requesting UP_PAGE (waittype 0x8007), blocked by: IX_PAGE: spid 18, dbid 6, page 0x2c5, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCA' where column1 = 100 and column2 = 'AAAA' spid 18 waiting for UP_PAGE (waittype 0x8007), blocked by: IX_PAGE: spid 17, dbid 6, page 0x2c5, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCB' where column1 = 105 and column2 = 'AAAB' VICTIM: spid 17, pstat 0x0000, cputime 20

Spid 17 (connection one) is waiting for an UP_PAGE lock, which is the first step to getting an exclusive page lock. It is being blocked by spid 18, which holds IX_PAGE lock on page 0x2c5. Spid 18 is waiting for UP_PAGE lock on the same page, and is being blocked by IX_PAGE lock held by spid 17. This leads to a deadlock because IX_PAGE lock is sharable, whereas UP_LOCK is not. During the first inserts, both the spids got IX_PAGE lock on the same page, and later they tried to upgrade the lock to UP_PAGE lock, which is not possible because UP_PAGE lock is exclusive.

The one way to avoid the deadlock is to insert the updated value directly into the table instead of inserting and then updating the row in the same transaction. If this is not possible, using the following command to disable IRL will help to avoid deadlock:

<pre class="fixed_text">  sp_tableoption 'example1', 'insert row lock', false

Example 4: Insertions to Rows on the Same Page
A deadlock may also result when the rows the two spids are working on are different but belong to the same page.

<pre class="fixed_text">  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC') Connection2 > INSERT INTO example1 VALUES (400, 'AAAB', 'CCC') Connection1 > UPDATE example1 SET column3 = 'CCCA' where column1 = 405 and column2 = 'AAAA'

At this point, Connection1 may be blocked by Connection2. This situation may occur because Connection1 wants to update a row in a page where Connection2 has already inserted a row.

<pre class="fixed_text">  Connection2 > UPDATE example1 SET column3 = 'CCCB' where column1 = 105 and column2 = 'AAAB'

At this point, Connection2 may also be blocked by Connection1, which will lead to a deadlock. This situation may occur when Connection2 wants to update a row in a page where Connection1 has inserted a row. The following is the deadlock trace:

<pre class="fixed_text">  97/04/20 15:48:21.18 spid20   *** DEADLOCK DETECTED with spid 19 *** spid 20 requesting UP_PAGE (waittype 0x8007), blocked by: EX_PAGE: spid 19, dbid 6, page 0x2c4, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCB' where column1 = 105 and column2 = 'AAAB' spid 19 waiting for UP_PAGE (waittype 0x8007), blocked by: EX_PAGE: spid 20, dbid 6, page 0xc48, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCA' where column1 = 405 and column2 = 'AAAA' VICTIM: spid 20, pstat 0x0000, cputime 60

This deadlock can be avoided by spreading out the rows over different pages. The one method to do this is to re-create the clustered index on this table with a large fill factor. The following is a statement that creates a clustered index with a fill factor of 50 percent:

<pre class="fixed_text">  create unique clustered index ex1ind1 on example1 (column1, column2) with fill factor = 50, PAD_INDEX

This statement creates the clustered index leaving half of the pages empty, including the non-leaf levels of the clustered index (because of the PAD_INDEX option). The table occupies double the actual size, and the number of rows per page are half of what they were.

The fill factor is not maintained on a table; the table is re-organized with the specified fill factor only during index creation time. Over time, the rows per page will change from the fill factor specified during index creation. When this occurs, it may be a good idea to re-create the clustered index with the desired fill factor.

Another solution to avoid the previous deadlock situation is to pad the table with dummy columns (for example, dummy1 char(255)). This increases the size of the row and leads to fewer rows per page (as few as one row per page). Because this type of padding is maintained over time, you do not need to re-create the clustered index to maintain the padding (though you may want to re-create the clustered index for other reasons). The disadvantage of this technique is that storage space is wasted on dummy fields.

Example 5: Padding Rows
Padding rows leads to fewer rows per page (hence fewer deadlocks), but it will not completely eliminate deadlocks.

In this example table, example1 is padded to occupy one row per page. The following are the statements used to create the table for this example:

<pre class="fixed_text">  create table example1 (column1 int, column2 char(20), column3 char(50),   dummy_column4 char (255), dummy_column5 char (255), dummy_column6 char   (255)) go  create unique index ex1ind5 on example1 (column3, column2, column1,   dummy_column4, dummy_column5, dummy_column6) with fill factor = 85 go  Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCC', ' ', ' ',  ' ', ' ') Connection2 > INSERT INTO example1 VALUES (400, 'AAAB', 'CCC', ' ', ' ',  ' ', ' ') Connection1 > UPDATE example1 SET column3 = 'CCCA' where column1 = 401 and column2 = 'AAAA'

At this point, Connection1 is blocked by Connection2 while updating the row. Because SQL Server must maintain page-chain pointers, it locks the previous page, the next page, and the page that is being updated. Because Connection2 holds a lock on the previous page, Connection1 must wait until Connection2 commits the transaction.

<pre class="fixed_text">  Connection2 > UPDATE example1 SET column3 = 'CCCB' where column1 = 101 and column2 = 'AAAB'

At this point, Connection2 is blocked by Connection1 because it must lock the previous page, which is currently locked by Connection1. The result is a deadlock. The following is the deadlock trace:

<pre class="fixed_text">  spid 20 requesting UP_PAGE (waittype 0x8007), blocked by: EX_PAGE: spid 19, dbid 6, page 0x12b5, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCB' where column1 = 101 and column2 = 'AAAB' spid 19 waiting for UP_PAGE (waittype 0x8007), blocked by: EX_PAGE: spid 20, dbid 6, page 0x1531, table example1, indid 0 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCA' where column1 = 401 and column2 = 'AAAA' VICTIM: spid 20, pstat 0x0000, cputime 300

This deadlock can be avoided by inserting dummy rows between the rows that are being inserted, updated, or deleted. For example, if Connection1 works (inserts, updates, or deletes) with row pk = 1 and Connection2 works with row pk = 5, inserting a row between these two rows (such as a row containing pk = 3) will avoid deadlocks. This method also increases the size of the table, but may be the best solution for those queue tables critical to the application.

Example 6: Nonclustered Indexes
In some cases, non-clustered secondary indexes may introduce deadlocks. In this example, the maintenance of the secondary index introduces deadlock.

The following is the statement used to create the secondary index used in this example:

<pre class="fixed_text">  create index ex1ind2 on example1 (column3) with fill factor = 90, PAD_INDEX Connection1 > BEGIN TRANSACTION Connection2 > BEGIN TRANSACTION Connection1 > INSERT INTO example1 VALUES (100, 'AAAA', 'CCBA', ' ', '  ', ' ', ' ') Connection2 > INSERT INTO example1 VALUES (300, 'AAAB', 'CCCZ', ' ', '  ', ' ', ' ') Connection2 > UPDATE example1 SET column3 = 'CCBA' where column1 = 105

At this point, Connection2 may be blocked by Connection1 because Connection1 may be holding a lock on the secondary non-clustered index page where Connection2 needs to update.

<pre class="fixed_text">  Connection1 > UPDATE example1 SET column3 = 'CCCZ' where column1 = 305

At this point, Connection1 may be blocked by Connection2, resulting in a deadlock. This situation can happen when Connection1 is waiting for a lock to update the non-clustered secondary index where Connection2 has already inserted and holds a lock on that page. The following is the deadlock trace for this deadlock example:

<pre class="fixed_text">  97/04/20 19:05:38.75 spid11   *** DEADLOCK DETECTED with spid 12 *** spid 11 requesting EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 12, dbid 6, page 0x112f, table example1, indid 0x2 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCCZ' where column1 = 305 spid 12 waiting for EX_PAGE (waittype 0x8005), blocked by: EX_PAGE: spid 11, dbid 6, page 0x1108, table example1, indid 0x2 pcurcmd UPDATE(0xc5), input buffer: UPDATE example1 SET column3 = 'CCBA' where column1 = 105 VICTIM: spid 11, pstat 0x0000, cputime 50

This deadlock can be avoided by dropping the secondary index. It is not possible to pad the index to contain one row per page, so this situation can be avoided only by eliminating the non-clustered secondary index or by modifying the application.

Deadlocks may occur with more than two connections, in which case the deadlock trace lists the spids involved in the deadlock and also the conflicting locks. Deadlocks may occur with RLOCK and XRLOCK locks, which are acquired during index traversing. Deadlocks may also occur because of extent locks (PR_EXT, NX_EXT, UPD_EXT & EX_EXT).

For additional information about analyzing deadlocks, you can enable the following trace flags:

T1200

Prints all of the lock request/release information when it occurs, whether a deadlock is involved or not. This is expensive in terms of performance, but it can be useful for analysis.

T1206

Prints all of the locks held by participating spids in the deadlock.

T1208

Prints the host name and program name supplied by the client. This can help identify a client involved in a deadlock, assuming the client specifies a unique value for each connection.

Keywords: kbhowto kbusage KB169960

-

[mailto:TECHNET@MICROSOFT.COM Send feedback to Microsoft]

© Microsoft Corporation. All rights reserved.