Cleanup prepared statement during node close

[marco6]

After preparing more queries on k8s-dqlite, I've been hitting quite frequently some segfaults. They all seemed to be happening when a query is running and the node is shutting down.

Apparently, they were. There was some missing cleanup logic around prepared statements in gateway__leader_close. This PR adds that bit.

I see that codecov report is showing a small regression in coverage. If someone from the dqlite team could guide me in testing this, I would be up for it.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 77.87%. Comparing base (054b511) to head (e6ac700).
Report is 8 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #675      +/-   ##
==========================================
+ Coverage   77.84%   77.87%   +0.02%     
==========================================
  Files         197      197              
  Lines       27761    27779      +18     
  Branches     5493     5538      +45     
==========================================
+ Hits        21611    21633      +22     
- Misses       4333     4420      +87     
+ Partials     1817     1726      -91     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[cole-miller]

I see that codecov report is showing a small regression in coverage. If someone from the dqlite team could guide me in testing this, I would be up for it.

I believe this is an artifact of codecov comparing the coverage numbers between your branch's HEAD and master's HEAD, instead of using the merge commit. If you look at the report, the patch lines that are listed as uncovered are all from the diff between merge_base(marco6/fix-prepared-cleanup, canonical/master), not from your commit. (I've noticed this happening before with other PRs.) So the coverage is not a blocker for merging.

[cole-miller]

For testing, I would try to add something in test/unit/test_gateway.c, which supports triggering gateway closure (grep for CLUSTER_DEPOSE). The tricky part will be getting that to happen while the gateway is in the middle of sending rows. Maybe this could be done relatively easily by just having a query that returns a ton of rows, even something like SELECT * FROM generate_series(1)? That might end up being slow or terribly CPU-intensive, though. If it doesn't work out, we could add some fault-injection that enables pausing the uv loop thread while it's in the middle of sending rows, so the window isn't so narrow. [Edit: ignore this, of course the gateway unit tests are single-threaded, so there is no "window" to hit.]

[cole-miller]

Test looks great, thank you!