Project

General

Profile

Bug #11846

Inconsistent session cache when evicting last record of a ProgressiveResults bracket

Added by Alexandru Lungu 15 days ago. Updated 15 days ago.

Status:
Review
Priority:
Normal
Assignee:
Alexandru Lungu
Target version:
-
Start date:
Due date:
% Done:

100%

billable:
No
vendor_id:
GCD
case_num:
version_reported:
version_resolved:
reviewer:
Constantin Asofiei
production:
No
env_name:
topics:

History

#2 Updated by Alexandru Lungu 15 days ago

  • % Done changed from 0 to 100
  • Assignee set to Alexandru Lungu
  • Status changed from New to Review
  • reviewer Constantin Asofiei added

I discovered this bug in my tests with two sessions (one reading and the other one writing) in a stressful manner. In ProgressiveResults.getResults:

         if (delegate != null && delegate.last())
         {
            lastRecord = delegate.get();
         }

lastRecord is a reference to a record that is cached.

After this, the delegate is closed and a new bracket is retrieved and eventually delegate.first() is called. There is a chance that this delegate.first() to require space in the Session.cache and actually evict lastRecord which is not pinned by any buffer.

Later on:

bufMgr.evictDMOIfUnused(persistence, null, dmo);

Evicting the dmo (which is computed from lastRecord) yields Inconsistent session cache log because we try to evict a DMO that is no longer cached. This is not fatal, but may trigger a severe log.

The real functional risk is that the eviction is triggered under !dmo.isInUse(), but this dmo may not be cached, so we end up running session.evict(dmo); and evict the cached image which might still be referenced.

I fixed this in 11846a/rev. 16746

Constantin, please review.

Also available in: Atom PDF