ryan@nardis.ca wrote:

On Fri, Apr 28, 2017 at 02:52:44PM +0000, ryan@nardis.ca wrote:

...

I tried cyrus-sasl-2.1.25 and the issue doesn't seem to happen. I'll see if I can isolate the related change.

SASL version was a red herring. I accidentally linked with Debian's libldap (2.4.31) when I tested that. Mea culpa.

The culprit does indeed seem to be the ITS#6798 change, as Quanah suggested. Reverting b483ee1a ("Drop ldap_int_sasl_mutex") on RE24 and linking with libldap_r makes my test program work.

Since the proof-of-concept patch I posted above also makes my test program work, I'm back to thinking that calling sasl_client_init concurrently is the real bug here.

...

On Fri, Apr 28, 2017 at 04:43:33AM +0000, ryan@openldap.org wrote:

...

If I'm right about this bug, I suppose the right fix is to wrap sasl_client_init in a new mutex.

... of course that idea's half-baked, because libldap doesn't *have* mutexes. hmm.

I'm still not happy with the idea of calling sasl_client_init from ldap_int_initialize, as it does a bunch of work (loading plugins and such) that non-SASL users don't care about. However, I haven't as yet thought of another fix that works for libldap as well as libldap_r.

Would love to hear others' thoughts on this. Maybe fixing for libldap_r only (and documenting that) is a good enough compromise?

Yes, if the problem only exists in concurrent use then it's only relevant to libldap_r.