[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: (ITS#6629) send_search_entry() does not clear controls



>
>
> --On August 26, 2010 3:44:02 AM +0000 masarati@aero.polimi.it wrote:
>
>> Full_Name: Pierangelo Masarati
>> Version: HEAD/re24
>> OS: irrelevant
>> URL: ftp://ftp.openldap.org/incoming/
>> Submission from: (NULL) (129.72.175.5)
>> Submitted by: ando
>>
>>
>> REP_CTRLS_MUSTBEFREED is only honored by send_ldap_response(); when
>> response controls are attached to slap_send_search_entry() it is not
>> honored.  This does not surface because controls must be malloc'ed on
>> the
>> slab; however, it clearly appears when built with SLAP_NO_SL_MALLOC.
>
> Interesting -- I'm seeing an issue where I can get slapd to consume all
> available memory on the system by using paged results.  It takes a while
> to
> do so (depending on how much memory is available), but it's 100%
> repeatable.  The more mixes of paged and non paged results there are with
> large result sets, the more quickly it happens.

pagedResults only sends a control response with the searchResultDone
response, which uses send_ldap_response() and thus clears the controls;
moreover, we are talking about few bytes, so I don't think this is the
issue.  There might be something else related to slab allocation that runs
out of slab memory and starts using the heap without freeing it.

The present issue might affect syncrepl, since in that case a control is
attached to each entry/intermediate response (afaik).

It might even be intentional; in fact, often when a control is attached to
an entry in a searchResultEntry the same control needs to be applied to
many entries returned by the same operation.  However, in this case we
should probably define a specific API for slap_add_ctrls() that does not
set the REP_CTRLS_MUSTBEFREED, or something like that.

> However, I do not build
> with SLAP_NO_SL_MALLOC (or at least, not intentionally ;) ).

I occasionally do (you only need to recompile slapd/sl_malloc.c) to trace
possible leaks when developing something.  Since we make an intensive use
of slab mallocs, usual minimal tests with valgrind do not show leaks
otherwise.

p.