From mboxrd@z Thu Jan  1 00:00:00 1970
X-Spam-Checker-Version: SpamAssassin 3.4.4 (2020-01-24) on polar.synack.me
X-Spam-Level: 
X-Spam-Status: No, score=-1.9 required=5.0 tests=BAYES_00 autolearn=ham
	autolearn_force=no version=3.4.4
X-Google-Language: ENGLISH,ASCII-7-bit
X-Google-Thread: 103376,9d303864ae4c70ad
X-Google-Attributes: gid103376,public
X-Google-ArrivalTime: 2004-04-16 15:15:46 PST
Path: archiver1.google.com!postnews1.google.com!not-for-mail
From: wojtek@power.com.pl (Wojtek Narczynski)
Newsgroups: comp.lang.ada
Subject: Re: Reprise: 'in out' parameters for functions
Date: 16 Apr 2004 15:15:46 -0700
Organization: http://groups.google.com
Message-ID: <5ad0dd8a.0404161415.220cce54@posting.google.com>
References: <u4qrv5hwr.fsf@acm.org> <bRecOT0TxF@VB1162.spb.edu>
 <mailman.224.1081468027.327.comp.lang.ada@ada-france.org>
 <5ad0dd8a.0404090512.15af2908@posting.google.com>
 <c56ib6$5ha$1@a1-hrz.uni-duisburg.de>
 <5ad0dd8a.0404091828.6e79bb4e@posting.google.com>
 <8Oadneu6eY9pweDdRVn-hA@comcast.com>
 <5ad0dd8a.0404151752.4f598e1a@posting.google.com>
 <VvadnXNuIcYs7OLdRVn-hg@comcast.com>
 <5ad0dd8a.0404160338.5a1c9116@posting.google.com>
 <1080a21p0nn482f@corp.supernews.com>
NNTP-Posting-Host: 83.27.43.192
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 8bit
X-Trace: posting.google.com 1082153746 21577 127.0.0.1 (16 Apr 2004 22:15:46
 GMT)
X-Complaints-To: groups-abuse@google.com
NNTP-Posting-Date: Fri, 16 Apr 2004 22:15:46 +0000 (UTC)
Xref: archiver1.google.com comp.lang.ada:7229
Date: 2004-04-16T15:15:46-07:00
List-Id: <comp.lang.ada>

Randy,

> That's because it would be hard to avoid deadlock in such data structures.

Yes, it is hard to avoid deadlocks.

> The usual way of avoiding deadlock is the use the onion-skin model of
> locking. 

With acyclic strucutres you do "crabbing" to avoid deadlocks.

> To see the problem, consider a tree struction. You'd have to lock many nodes
> if an insertion caused a rebalancing. If a second task tried to do a nearby
> insertion at the same time, it also could lock a number nodes. It would be
> quite likely for the second task to lock the nodes that it is going to
> insert into, then the first task try to lock the same nodes to do a
> rebalancing. So it would have to wait. Then the second task would discover
> that it, too needs to do a rebalancing - but it can't, because the first
> task already has locked some of the nodes that it needs to change for the
> rebalancing. Classic deadlock. To avoid this situation, you have to lock the
> whole data structure. 

This is not easy but solveable. For example ADABAS (and forks)
relational database is organized as one large B* tree. It can be
edited concurrently, it rebalances, and does not deadlock.

One more example: Java heap, millions of objects, 64 CPUs, thousands
of threads. Would you like to have one lock for the whole data
strucutre?

Regards,
Wojtek