Ludovic Poitou wrote:
Sun Directory Server, OpenDS, OpenDJ, Port 389 (formerly known as Fedora
Directory) at least, do not support Alias dereferencing because the complexity
out-weights the benefits, and very few client applications make use of it
anyway.
My 2 cents.
We had no alias support in back-bdb originally; I added it back in 2003.
Indeed, you only need to track all of the alias entries that you hit, and
that's typically a small number. Each new entry defines a new search scope.
Any time a dereference leads to an entry that resides anywhere within your
already-known search scopes that deref can be considered fully resolved,
because the result was already part of the scope.
It's actually quite simple and quite fast. Using the objectclass index it's
trivial to obtain the list of all alias entries within the database, so from
the outset you already know the maximum size of what you're dealing with.
And just to remind, alias deref is a 2-bit option - deref base of the search,
and deref (the rest of the search). They're selected independently.
Ludo
---
Ludovic Poitou
http://ludopoitou.wordpress.com
On Mon, Jun 13, 2011 at 4:39 PM, Emmanuel Lecharny <[email protected]
<mailto:[email protected]>> wrote:
Alias cycle detection
---------------------
There is an unsolved question about how we should detect Alias cycles.
Right now, we check for cycles *before* they are created. The alternative
would be to stop any search that could lead to an infinite loop.
A third - but unrealistic - solution would be to don't detect cycle, and
process the search until we reach the time or size limit (in other words,
it's up to the admin to avoid the creation of such cycle; Highly
dangerous…).
The problem with the first approach is that we can't anymore pass the
VSLDAP tests. It's a major burden. Also most of the current servers
support this feature.
So it seems that everything concurs to get the cycle be detected while
searching, not before. Now, how can we do that efficiently ?
Detecting a cycle is a pretty obvious algorithm : every entry we went
through should be memorized, and we should check that we aren't going
through an entry we already returned.
This is easy, but highly memory consuming, as we have to keep track of
*all* the entries during a search. Simply idealistic.
Hopefully, we can use a simplified algorithm : if we hit an Alias, we have
to check that we don't come back to the original DN we came from, or one
of its descendant. We should also memorize each indirection (each of them
is like a new search)
initial search : A.B.C -~-> a(alias) (A.B.C memorized)
a -~-> B.C… -~-> A.B.C : cycle detected (B.C memorized)
or
initial search : A.B.C -~-> a(alias) (A.B.C memorized)
a -~-> D.E… -~-> b(alias) : (D.E memorized)
b -~-> B.C… -~-> A.B.C : cycle detected (B.C memorized)
This way we only have to memorize the roots for all searches, instead of
memorize all he entries. As we are very unlikely to have many aliases and
many indirection, it should be safe from the memory consumption POV.
One other aspect is that this should be only done if the user has required
that we dereference aliases on the server.
Last, not least, if the user has sent the ManageDSAIt control, we should
not have to memorize anything either, as we don't dereference aliases in
this case.
The DN cache will be stored in the session, and removed as soon as we
reached the end of the search. It will be associated with the message ID,
so that a user can issue more than one search in parallel.
thoughts ?
--
Regards,
Cordialement,
Emmanuel Lécharny
www.iktek.com <http://www.iktek.com>
--
-- Howard Chu
CTO, Symas Corp. http://www.symas.com
Director, Highland Sun http://highlandsun.com/hyc/
Chief Architect, OpenLDAP http://www.openldap.org/project/
