Use write barrier when setting busy_handler

[jhawthorn]

The Database type struct is marked as WB_PROTECTED which means that any new references added to the object need to fire the write barrier. Previously we were missing this when setting the busy_handler.

I wasn't able to make a smaller reproduction, but I started investigating this after seeing this crash in the Rails CI suite: https://buildkite.com/rails/rails/builds/111147#0191d7ec-de81-49c8-b20a-d207f446d45e

<OBJ_INFO:[email protected]:7072> 0x00007f4fd042cb20 [0 M    ] T_NONE
test/cases/connection_pool_test.rb:1041: [BUG] try to mark T_NONE object
ruby 3.3.5 (2024-09-03 revision ef084cc8f4) [x86_64-linux]
-- Control frame information -----------------------------------------------
c:0060 p:---- s:0311 e:000310 CFUNC  :initialize
c:0059 p:---- s:0308 e:000307 CFUNC  :new
c:0058 p:0012 s:0304 e:000303 METHOD test/cases/connection_pool_test.rb:1041
-----8<-----
-- C level backtrace information -------------------------------------------
/usr/local/lib/libruby.so.3.3(rb_print_backtrace+0x14) [0x7f4fecfcfaeb] /usr/src/ruby/vm_dump.c:820
/usr/local/lib/libruby.so.3.3(rb_vm_bugreport) /usr/src/ruby/vm_dump.c:1151
/usr/local/lib/libruby.so.3.3(bug_report_end+0x0) [0x7f4fecdc8a7a] /usr/src/ruby/error.c:1042
/usr/local/lib/libruby.so.3.3(rb_bug_without_die) /usr/src/ruby/error.c:1042
/usr/local/lib/libruby.so.3.3(die+0x0) [0x7f4fecd064ab] /usr/src/ruby/error.c:1050
/usr/local/lib/libruby.so.3.3(rb_bug) /usr/src/ruby/error.c:1052
/usr/local/lib/libruby.so.3.3(gc_mark_ptr+0x1ad) [0x7f4fecdeff4d] /usr/src/ruby/gc.c:7073
/usr/local/bundle/gems/sqlite3-2.0.4-x86_64-linux-gnu/lib/sqlite3/3.3/sqlite3_native.so(0x7f4fd01d55d6) [0x7f4fd01d55d6]
/usr/local/lib/libruby.so.3.3(gc_mark_stacked_objects+0x78) [0x7f4fecdf3408] /usr/src/ruby/gc.c:7565
/usr/local/lib/libruby.so.3.3(gc_mark_stacked_objects_all) /usr/src/ruby/gc.c:7603
/usr/local/lib/libruby.so.3.3(gc_marks_rest) /usr/src/ruby/gc.c:8798
-----8<-----

Given that afaict this is the only mark function within sqlite3 I think it's pretty likely this was the cause.

[flavorjones]

@jhawthorn Thanks for this!

[flavorjones]

I'll cut a release in the next day or so.

[flavorjones]

Updated the changelog in 6c274b4, feel free to suggest alternate wording.

[jhawthorn]

That wording sounds fine, although we could be less specific about how the crash will surface (I don't think it will always be in the mark phase, we could also be super unlucky and end up with a different object in place of the busy handler).

When setting a Database busy_handler, fire the write barrier to prevent potential crashes on Garbage Collection.