15:55 < jhrozek> What I'm seeing is that sssd_nss always returns *all* members, returns all to sss_nss, which can't stuff the results into the given buffer and returns TRYAGAIN.
15:56 < simo> jhrozek, ok then we need an optimization where we cache the last call
15:56 < simo> so that repeated calls due to small buffers do not cause roundtrips
15:56 < sgallagh> simo: Except, what happens if glibc does not in fact TRYAGAIN?
15:56 < simo> jhrozek, it should be a very simple patch
15:56 < simo> jhrozek, we already do that for enumerations
15:57 < simo> sgallagh, the next get** call will free the cache
15:57 < sgallagh> Ok, true.
15:57 < simo> or we will waste 8k of memory ... big deal
15:57 < sgallagh> Sounds like a plan, then
15:57 < simo> the cache needs to be a global
15:57 < simo> and needs locking for multithreaded applications
15:57 < simo> uhmm maybe we can avoid locking and have just thread-local storage for this one ...
15:58 < simo> a mutex is simpler for now (we already have a mutex around the fd communication
15:58 < jhrozek> simo: sounds good to me. Would you like to file a ticket or shall I?
15:59 < simo> and will keep memory low as only one storage will be involved, with tls we could have hundreds of abandoned buffers
15:59 < simo> jhrozek, feel free
15:59 < simo> ah one last thing
15:59 < simo> cache *only* if the buffer is too small
15:59 < simo> no need for caching in libnss_sss otherwise we would just thrash constantly
15:59 < jhrozek> OK