Add extension point to record execution times in a performance log

Add a new extension point that is invoked for all operations for which
the execution time is measured. The invocation of the extension point
does not happen immediately, but only at the end of a request (REST
call, SSH call, git push). Implementors can write the execution times
into a performance log for further analysis.

In Gerrit there are 2 possibilities to measure and record the execution
time of an operation:

1. TraceTimer:
   Opens an autocloseable context to execute an operation. On close the
   execution time is written to the server logs, but only if the request
   is traced (or if log level was set to fine).
2. Timer Metrics:
   Record execution times as a metric.
   In addition this writes the execution time to the server logs, but
   only if the request is traced (or if log level was set to fine).

These are the 2 places where performance log entries must be captured.

Performance log entries are stored in the LoggingContext which is based
on ThreadLocals. LoggingContextAwareRunnable and
LoggingContextAwareCallable which are used by all executors ensure that
captured performance log entries are properly copied between threads.

At the end of a request (supported are REST call, SSH call and git push)
the captured performance log entries are handed over to the
PerformanceLogger implementations.

If no PerformanceLogger is registered, or if execution times of
operations outside of request scope are measured, performance log
entries are not captured because nobody would consume them.

Storing the performance log records in the LoggingContext and invoking
the PerformanceLogger plugins only at the end of a request has some
advantages and disadvantages:

1. [advantage] Users of TraceTimer can continue to use the static
   methods calls to create a TraceTimer (TraceContext.newTimer(...)). To
   invoke the plugins immediately we would need to get them injected
   into TraceTimer, hence callers would need a factory to create a
   TraceTimer. That would result in quite some boilerplate code and in
   addition some places that use TraceTimer (in VersionedMetaData)
   cannot use injection.
2. [advantage] The metric system is setup very early in the injector
   chain (in the DB injector, see SiteProgram#createDbInjector(boolean)).
   To invoke the plugins directly from the timer metrics we would need
   to have the PerformanceLogger plugins already available on this
   injector level, which is difficult.
3. [disadvantage] The captured performance log records are kept in
   memory while a request is processed. This leads to higher memory
   foodprint, but we think this is OK.

To keep the performance and memory overhead for recording performance as
low as possible we did some optimizations:

1. Performance log entries are only created if there is a consumer (at
   least one PerformanceLogger plugin + time measured inside request
   context)
2. Performance log entries avoid the instantiation of a Map to record
   meta data (instead we have dedicated fields for meta data keys and
   values).

For the timer metrics we use generic names for the meta data keys
("field1", "field2", "field3"). This is because the actual field names
are not available at this place. We may make them available, but that's
outside the scope of this change and may be done in a follow-up change.

To be able to write stable acceptance tests that verify that the
PerformanceLogger plugins are invoked, it is important that the server
calls the plugins before the response is sent back to the client. This
is why the scope of the PerformanceLogContext in RestApiServlet is a
little smaller than the scope of the TraceContext.

Change-Id: I699db01609a1b4a88cee8959bdd9f1dfbb8dc74e
Signed-off-by: Edwin Kempin <ekempin@google.com>

java/com/google/gerrit/metrics/Timer0.java

@@ -18,6 +18,8 @@ import static java.util.concurrent.TimeUnit.NANOSECONDS;
 
 import com.google.common.flogger.FluentLogger;
 import com.google.gerrit.extensions.registration.RegistrationHandle;
+import com.google.gerrit.server.logging.LoggingContext;
+import com.google.gerrit.server.logging.PerformanceLogRecord;
 import java.util.concurrent.TimeUnit;
 
 /**
@@ -68,7 +70,10 @@ public abstract class Timer0 implements RegistrationHandle {
    * @param unit time unit of the value
    */
   public final void record(long value, TimeUnit unit) {
-    logger.atFinest().log("%s took %dms", name, unit.toMillis(value));
+    long durationMs = unit.toMillis(value);
+    LoggingContext.getInstance()
+        .addPerformanceLogRecord(() -> PerformanceLogRecord.create(name, durationMs));
+    logger.atFinest().log("%s took %dms", name, durationMs);
     doRecord(value, unit);
   }