Re: some question about deadlock

From: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
To: "ipig" <ipig(at)ercist(dot)iscas(dot)ac(dot)cn>
Cc: "Bruce Momjian" <pgman(at)candle(dot)pha(dot)pa(dot)us>, pgsql-hackers(at)postgresql(dot)org
Subject: Re: some question about deadlock
Date: 2006-05-29 15:51:38
Message-ID: 14600.1148917898@sss.pgh.pa.us
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

"ipig" <ipig(at)ercist(dot)iscas(dot)ac(dot)cn> writes:
> That is to say, if p0 wants to lock A again, then p0 will be put before p1, and p0 will be at the head of the queue. Why do we need to find the first waiter which conflicts p0? I think that p0 must be added at the head of the wait queue.

Your analysis is assuming that there are only two kinds of lock, which
is not so. Process A might hold a weak lock and process B a slightly
stronger lock that doesn't conflict with A's. In the wait queue there
might be process C wanting a lock that conflicts with B's but not A's,
followed by process D wanting a strong lock that conflicts with all three.
Now suppose A wants to get a lock of the same type D wants. Since this
conflicts with B's existing lock, A must wait. A must go into the queue
before D (else deadlock) but if possible it should go after C, on
fairness grounds.

A concrete example here is
A has AccessShareLock (reader's lock)
B has RowExclusiveLock (writer's lock)
C wants ShareLock (hence blocked by B but not A)
D wants AccessExclusiveLock (must wait for all three)
If A wants to upgrade to AccessExclusiveLock, it *must* queue in front
of D, and we'd prefer that it queue behind C not in front of C.

regards, tom lane

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Tom Lane 2006-05-29 16:01:09 Re: question about security hole CVE-2006-2313 and UTF-8
Previous Message Bruce Momjian 2006-05-29 15:43:29 Re: some question about deadlock