You seem to have found a number of problems with the new code that tries
to implement the unique constraint with multiple nulls. I definitely
missed the deleted row case when reviewing that code, and had assumed it
just ever needed to search one row left and right.
I am now wondering if it would be better to make this code work the same
as the existing code in the regular unique case. That code makes sure
there is always only ONE row with the given key (not including the row
id). It manipulates the search params so that either that exact row
is found or it positions in the right place to do the insert.
It does have some tricky code for when it finds a matching row,
especially if it is marked deleted, but that code has been working for
a long time now.
The approach would be something like:
o If there is a null in the key then just do the insert normally into
the non-unique table, using all keys.
o If no null, then need to change the search params to only use leading
columns and to not allow exact match - this means changing the existing
code which would use the
tables information which would say to use all columns.
o and then if match use the existing code for unique tables in this case.
Let me know what you think. Is this a better approach or should we
patch up the new code? Are you interested in doing this? If not I
may have time to do this.
/mikem
Knut Anders Hatlen (JIRA) wrote:
[
https://issues.apache.org/jira/browse/DERBY-4080?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Knut Anders Hatlen updated DERBY-4080:
--------------------------------------
Attachment: repro.sql
Here's a script that exposes the bug (seen in 10.4.2.1 - (706043) and 10.5.0.0
alpha - (749659)).
The script inserts a row at the end of one index page, with an uncommitted
deleted duplicate at the beginning of the next index page. The insert must wait
for the uncommitted delete to be committed. The transaction that holds the
exclusive lock on the deleted duplicate then tries to read a row on the
previous index page, but cannot obtain the lock until the insert operation
times out.
A thread dump during the hang shows that this is a deadlock involving both
locks and latches:
"Thread-2" prio=3 tid=0x08473800 nid=0x12 in Object.wait()
[0xb60de000..0xb60debe0]
java.lang.Thread.State: TIMED_WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
- waiting on <0xf3c06298> (a
org.apache.derby.impl.services.locks.ActiveLock)
at org.apache.derby.impl.services.locks.ActiveLock.waitForGrant(Unknown
Source)
- locked <0xf3c06298> (a
org.apache.derby.impl.services.locks.ActiveLock)
at
org.apache.derby.impl.services.locks.ConcurrentLockSet.lockObject(Unknown
Source)
.
.
.
"main" prio=3 tid=0x0806f400 nid=0x2 in Object.wait() [0xfe34e000..0xfe34ed38]
java.lang.Thread.State: WAITING (on object monitor)
at java.lang.Object.wait(Native Method)
- waiting on <0xf4235668> (a
org.apache.derby.impl.store.raw.data.StoredPage)
at java.lang.Object.wait(Object.java:485)
at org.apache.derby.impl.store.raw.data.BasePage.setExclusive(Unknown
Source)
- locked <0xf4235668> (a
org.apache.derby.impl.store.raw.data.StoredPage)
at org.apache.derby.impl.store.raw.data.BaseContainer.latchPage(Unknown
Source)
If the transaction that waits for the row lock had released all latches once it detected
that it would have to wait, there would not be a deadlock and both transactions would be
able to complete successfully. Once it has obtained the lock, it will release the latch
on the page that it "forgot" to unlatch and perform a rescan, so I would
believe that it is fine to release the latch earlier in this case.
Possible deadlock between locks and latches in
BTreeController.compareRowsForInsert()
-------------------------------------------------------------------------------------
Key: DERBY-4080
URL: https://issues.apache.org/jira/browse/DERBY-4080
Project: Derby
Issue Type: Bug
Components: Store
Affects Versions: 10.4.2.0
Reporter: Knut Anders Hatlen
Attachments: repro.sql
It looks like BTreeController.compareRowsForInsert(), which is used to check
for duplicates in a unique nullable index, can run into a deadlock which
involves both locks and latches.
Here's what I think can happen:
comparePreviousRecord() (or compareNextRecord()) holds a latch on the index
page where a new row is about to be inserted, and needs to check if there's a
duplicate on one of the adjacent rows. Because the row is near a page boundary,
this check moves to another index page, while still holding the latch on the
original index page. Then compareRowsForInsert() is called, which tries to get
an update lock on the existing row. If it has to wait for the update lock, the
latch on the current page is released, but the latch on the original index page
is kept. This means that the transaction is holding a latch while it is waiting
for a lock, which means that it is blocking all access to that page until it
has been granted the lock. If some other transaction that is holding a
conflicting lock on the row later needs to latch the index page, those two
transactions run into a deadlock and the one that's waiting for the lock will
eventually time out (but it will not be reported as a de
ad
lock).
If compareRowsForInsert() releases all latches when it needs to wait for a
lock, the deadlock is prevented, and both of the transactions may be able to
complete without timing out.
