On 05/29/2014 06:13 PM, michael@stroeder.com wrote:

Howard Chu wrote:

...

michael@stroeder.com wrote:

...

It seems that modify requests which failed due to Invalid DN syntax (34) are not written to accesslog-DB. I guess that those requests get abandoned by the frontend and never reach the backend at all.

Correct.

...

It would be handy to see the invalid modify request in the accesslog-DB though.

Any chance to achieve this?

Not likely. The frontend must call select_backend() based on the incoming DN to determine which backend to invoke, and thus which stack of overlays are involved. If the DN is invalid, no selection can occur.

Hmm, I've done some more tests. Invalid syntax (21) also does not make it beyond the frontend into accesslog-DB.

I have no clear opinion on this. Of course the current behaviour is good for performance. But sometimes one would like to observe what broken LDAP clients sent in a modify request in the past.

Also running with BER loglevel or breaking up the TLS connection with stunnel and sniff with Wireshark is not always an option.

Having this configurable would be great.

What's your opinion on this?

In terms of code architecture, overlays have been designed to work on validated data (the request must be well formed). To this end, the notion of "frontend" database was added to make it possible to layer an overlay *before* backend selection.

To have a custom piece of code muck with a request before it is validated (or even before it is parsed) you need to split the code into:

- parsing, without doing any sanity check

- allow custom code to step in

- do validation afterwards

This does not necessarily imply performance penalty (when no custom code steps in, the amount of operations would be the same for sane requests; in case of malformed requests, you would miss the opportunity to bail out on the first inconsistency).

However, it would imply significant refactorization of the current code, without (IMHO) a strong motivation for it: malformed requests are already handled in terms of response to clients, and somehow logged.

An (expensive) alternative would be to register a handler at connection level which duplicates the parsing to anticipate checking for errors that would be detected before passing control to the frontend. Not the way I'd go, unless strictly necessary.

A paranoid but intriguing alternative would be to implement an operation mode that keeps parsing after the first error is detected, turning the operation into a no-op. Again, lots of refactorization and error handling and so on (for example, what if a syntax, an attribute description or a matching rule is not found? You'd always have to check that the corresponding pointer is not NULL before using it, whereas right now we often assume that if execution got there, data must be consistent. Of course, we could introduce dummy syntaxes "NonExistingSyntax", ..., but we'd probably need to define their semantics in such a way that they apply to every case we need to handle and so).

My 2c.

p.