Invalid access or accessor meta-issue

[finnschiermer]

This meta issue track other issues that indicate use of a stale accessor.

There may be a single or multiple underlying root causes. We don't know.

We have 4 different ways such an error can manifest itself

• A no such table exception (all of these are now considered fixed - but leaving metadata here for now)
• A stale object reference exception (this is the same as above, just a name change)
• A failure that appears to be related to encryption or the ref translation machinery, but encryption is reported to not be used.
• A segfault or access error at a transaction boundary

Guesses re root cause:

• Abuse of the Core API by ObjectStore. A bug which produces (or merely allows) use of a stale accessor
• A bug in the way we track ref-translation changes and force accessor updates.
• A bug in the management of memory mappings causing a mapping to be removed too early.

Links (to be added):

• NoSuchTable
  • NoSuchTable: No such table exists realm-java#7100 (presumably fixed)
  • Crash due to a NoSuchTable exception realm-swift#6767 (likely fixed by Fix a race condition in notifier initialization #4234)
  • NoSuchTable thrown from import_copy_of() #3761 (likely fixed by Fix a race condition in notifier initialization #4234)
• Crashes in encryption layer, where encryption is not used
  • Realm 7.0.8 native crash - EncryptedFileMapping::get_local_index_of_address realm-java#7145
  • Realm (7.0.8) crashes when iterates through RealmResults (Key not found) realm-java#7174
  • realm::util::EncryptedFileMapping::read_barrier realm-swift#6763
• (internal) https://ci.realm.io/blue/organizations/jenkins/realm%2Frealm-js/detail/PR-3355/10/pipeline/856/

Historic note:

Prior to Core-6, we had a trickle of error reports pertaining to invalid access to the accessor tracking machinery. We never nailed the root cause of this, and the whole machinery was designed away for Core-6. If the root cause actually resided outside (above) Core in the software stack, it may still be there, just showing itself differently.

[finnschiermer]

I no longer believe it can be a bug in the memory mapping management. To stress this part of the machinery, a special version of Core was built, in which the section size was lowered to 16 kilobytes. This increases turnaround of memory mappings by several orders of magnitude, but requires special filtering of tests - since many of our tests cannot run with such a change. Tests which failed as expected where filtered out. No tests failed in unexpected ways (including anything mimicking the reported bugs). Multi process and multi threaded tests were then expanded to run for hours. We were still unable to trigger a crash.

[sync-by-unito]

➤ Finn Andersen commented:

The ref-translation bugs are all explained by bugs in selecting the proper allocator and thus indirectly getting to a stale ref-translation table.

This was fixed in multiple rounds, last in #4417. No ref-translation error has been reported since.

[sync-by-unito]

➤ Finn Andersen commented:

All errors in this meta-issue are now considered fixed.