Race condition on all SetOptions on "block_based_table_factory"

[pdillinger]

Apply this patch to expand SetOptions testing in stress test and amplify bad behavior:

diff --git a/db_stress_tool/db_stress_test_base.cc b/db_stress_tool/db_stress_test_base.cc
index 0c8b7ee08..b1200c3e5 100644
--- a/db_stress_tool/db_stress_test_base.cc
+++ b/db_stress_tool/db_stress_test_base.cc
@@ -254,6 +254,11 @@ bool StressTest::BuildOptionsTable() {
            "2",
        }},
       {"max_sequential_skip_in_iterations", {"4", "8", "12"}},
+      // Either of these suffices for TSAN race condition report
+      //{"block_based_table_factory", {"{block_cache=2M}",
+      //                               "{block_cache=4M}"}},
+      {"block_based_table_factory", {"{block_size=2048}",
+                                     "{block_size=4096}"}},
   };
 
   if (FLAGS_allow_setting_blob_options_dynamically) {
diff --git a/tools/db_crashtest.py b/tools/db_crashtest.py
index 005cda1ce..16b0b1e05 100644
--- a/tools/db_crashtest.py
+++ b/tools/db_crashtest.py
@@ -252,7 +252,7 @@ blackbox_default_params = {
     # since we will be killing anyway, use large value for ops_per_thread
     "ops_per_thread": 100000000,
     "reopen": 0,
-    "set_options_one_in": 10000,
+    "set_options_one_in": 100,
 }
 
 whitebox_default_params = {
@@ -285,7 +285,7 @@ simple_default_params = {
 
 blackbox_simple_default_params = {
     "open_files": -1,
-    "set_options_one_in": 0,
+    "set_options_one_in": 100,
 }
 
 whitebox_simple_default_params = {}

Run USE_CLANG=1 COMPILE_WITH_TSAN=1 make blackbox_crash_test and quickly observe race condition reported.

[pdillinger]

@mrambacher Is this something you can look into? The race condition seems rather "baked in" to the Configurable interface with GetOptionsPtr.

[mrambacher]

@pdillinger I will see if I can reproduce this on my laptop. This may be related to the IsMutable #8928, which attempted to avoid some of these race conditions, but that is just conjecture at this point.

[pdillinger]

As noted in #11929, after that change there seems to be only a read-write race (not a write-write race), which should be "generally OK / safe" for non-pointer, 64-bit-or-less options.