Hypothesis
The issue seems to be that there is a race condition between exqlite_close and exqlite_prepare. The issue happens when we execute sqlite3_prepare_v3 on a connection and in-parallel we close the connection with sqlite3_close_v2.
sqlite3_close_v2 does handle previously prepared statements, but here we are preparing the statement in-parallel while we are closing the connection. Which is causing occasional SQLITE_MISUSE.
How can we reach this state?
Given
disconnect callback is executed by connection processhandle_prepare and other callbacks are executed in the client process- a NIF call can not be interrupted/cancelled once started, irrespective of client side process timeout, or process kill etc.
Scenario
- client checkout a connection and makes a query
- client query function call invocation makes
exqlite_prepare NIF call but fail to complete within connection timeout duration. It is 1ms in the test case we are using. This NIF call is still running or scheduled to run even after the timeout.
- due to connection timeout, connection-process receives timeout and initiate connection close (
disconnect callback)
- connection-process calls
exqlite_close NIF
2 & 4 are run in-parallel, leading to race-condition mentioned previously, causing segfault.
Fix
https://github.com/elixir-sqlite/exqlite/pull/218
I tried to fix the race-condition using mutex, and sofar haven’t seen a single failure.
This is PoC, I guess we should handle multiple other cases as well, not just the one between close & prepare if this is the actual issue.
