http://bugzilla.opensuse.org/show_bug.cgi?id=1171128 http://bugzilla.opensuse.org/show_bug.cgi?id=1171128#c1 William Brown changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |william.brown@suse.com --- Comment #1 from William Brown --- Hi there, This has been detailed in https://bugzilla.suse.com/show_bug.cgi?id=1089640 which I don't believe you have access to. The summary is that a customer reported a crash in 2018, and appears to be difficult to reproduce. This patch was able to prevent some issues during the shutdown process, but further analysis points to the fact that non-atomic int handling is used between threads during openldap shutdown which causes the issue where workers may be able to reactivate and prevent proper resource de-allocation. As this only affected our customer, and in a reasonably old version, we did not feel it was worth contacting upstream about this matter as you would likely have not been interested in the effort required to resolve this. Of course, this being the open source branch of openldap, and being open to the public, if you have a concern with the patch, you are welcome to remove it or perform your own analysis and contributions, which we would happily review. So I will leave it to you to decide what action you feel is needed to improve this package, especially given you are far more knowledgeable and invested in openldap than many others :) Thanks, -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1171128 http://bugzilla.opensuse.org/show_bug.cgi?id=1171128#c3 William Brown changed: What |Removed |Added ---------------------------------------------------------------------------- Status|NEW |RESOLVED Resolution|--- |WORKSFORME --- Comment #3 from William Brown --- I can not attest to the original authors thoughts, but their comment on your ITS is: "After some investigation I narrowed the problem down to mdb_env_reader_dest destructor function being called after its object had been deleted while powering off the application.". It was also determined in the private bug that due to non-atomic int usage in openldap shutdown it's theoretically viable for a reader to awaken during shutdown and attempt to access mdb resources. So the origin of the issue is likely in incorrect use of non atomic ints between threads, and while the patch may appear "invalid", it was able to improve the customers experience during shutdowns. Given this information, I'm erring to retain the patch as it's presence is not harmful and there is no evidence that it causes issues or corruption. Even from a programming viewpoint, even if that variable were never access again, it is responsible defensive program to move the value to a state that prevents this kind of issue. -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1171128 http://bugzilla.opensuse.org/show_bug.cgi?id=1171128#c5 --- Comment #5 from William Brown --- That is not correct. No extra patches to affect shutdown have been added. You have threads created with: ldap_pvt_thread_create(..., ldap_int_thread_pool_wrapper, ...) On shutdown you have the following loop: while (pq->ltp_open_count) { ldap_pvt_thread_cond_broadcast(&pq->ltp_cond); ldap_pvt_thread_cond_wait(&pq->ltp_cond, &pq->ltp_mutex); } And the ldap_int_thread_pool_wrapper performs the decrement on shutdown: pq->ltp_open_count--; This is from your git master. These integer operations are non-atomic, and un-fenced, meaning that it is possible (and likely) that there may be disagreements between ltp_open_count between your threads. It appears also that ltp_active_count and ltp_finishing, and many other variables are also non-atomic, meaning that this thread pool is unsound and may have effects on systems at scale during shutdown that are non-deterministic. I think it's beyond my ability to resolve this in openldap as I am not familiar enough with the code base, but it is from my view very plausible that this is the cause of the issues. As a result, the patch stays. -- You are receiving this mail because: You are on the CC list for the bug.

http://bugzilla.opensuse.org/show_bug.cgi?id=1171128 http://bugzilla.opensuse.org/show_bug.cgi?id=1171128#c10 --- Comment #10 from William Brown --- (In reply to Michael Ströder from comment #9)

...

This has been detailed in https://bugzilla.suse.com/show_bug.cgi?id=1089640 which I don't believe you have access to. The summary is that a customer reported a crash in 2018, and appears to be difficult to reproduce. This patch was able to prevent some issues during the shutdown process, but further analysis points to the fact that non-atomic int handling is used between threads during openldap shutdown which causes the issue where workers may be able to reactivate and prevent proper resource de-allocation.

Simon Lees formerly provided bug information to me. IIRC there was no test system available at the customer's site to even reproduce the issue or to prove that this patch really fixes the customer's issue.

This was the point where I started the fork of my own openldap2 package.

You are more than welcome to take over the network:ldap version of openldap2 if you want. -- You are receiving this mail because: You are on the CC list for the bug.