Merge tag 'v0.33.3'

Features
--------

- Add support for the SNI extension to federation TLS connections. Thanks to @vojeroen! ([\#3439](https://github.com/matrix-org/synapse/issues/3439))
- Add /_media/r0/config ([\#3184](https://github.com/matrix-org/synapse/issues/3184))
- speed up /members API and add `at` and `membership` params as per MSC1227 ([\#3568](https://github.com/matrix-org/synapse/issues/3568))
- implement `summary` block in /sync response as per MSC688 ([\#3574](https://github.com/matrix-org/synapse/issues/3574))
- Add lazy-loading support to /messages as per MSC1227 ([\#3589](https://github.com/matrix-org/synapse/issues/3589))
- Add ability to limit number of monthly active users on the server ([\#3633](https://github.com/matrix-org/synapse/issues/3633))
- Support more federation endpoints on workers ([\#3653](https://github.com/matrix-org/synapse/issues/3653))
- Basic support for room versioning ([\#3654](https://github.com/matrix-org/synapse/issues/3654))
- Ability to disable client/server Synapse via conf toggle ([\#3655](https://github.com/matrix-org/synapse/issues/3655))
- Ability to whitelist specific threepids against monthly active user limiting ([\#3662](https://github.com/matrix-org/synapse/issues/3662))
- Add some metrics for the appservice and federation event sending loops ([\#3664](https://github.com/matrix-org/synapse/issues/3664))
- Where server is disabled, block ability for locked out users to read new messages ([\#3670](https://github.com/matrix-org/synapse/issues/3670))
- set admin uri via config, to be used in error messages where the user should contact the administrator ([\#3687](https://github.com/matrix-org/synapse/issues/3687))
- Synapse's presence functionality can now be disabled with the "use_presence" configuration option. ([\#3694](https://github.com/matrix-org/synapse/issues/3694))
- For resource limit blocked users, prevent writing into rooms ([\#3708](https://github.com/matrix-org/synapse/issues/3708))

Bugfixes
--------

- Fix occasional glitches in the synapse_event_persisted_position metric ([\#3658](https://github.com/matrix-org/synapse/issues/3658))
- Fix bug on deleting 3pid when using identity servers that don't support unbind API ([\#3661](https://github.com/matrix-org/synapse/issues/3661))
- Make the tests pass on Twisted < 18.7.0 ([\#3676](https://github.com/matrix-org/synapse/issues/3676))
- Don’t ship recaptcha_ajax.js, use it directly from Google ([\#3677](https://github.com/matrix-org/synapse/issues/3677))
- Fixes test_reap_monthly_active_users so it passes under postgres ([\#3681](https://github.com/matrix-org/synapse/issues/3681))
- Fix mau blocking calulation bug on login ([\#3689](https://github.com/matrix-org/synapse/issues/3689))
- Fix missing yield in synapse.storage.monthly_active_users.initialise_reserved_users ([\#3692](https://github.com/matrix-org/synapse/issues/3692))
- Improve HTTP request logging to include all requests ([\#3700](https://github.com/matrix-org/synapse/issues/3700))
- Avoid timing out requests while we are streaming back the response ([\#3701](https://github.com/matrix-org/synapse/issues/3701))
- Support more federation endpoints on workers ([\#3705](https://github.com/matrix-org/synapse/issues/3705), [\#3713](https://github.com/matrix-org/synapse/issues/3713))
- Fix "Starting db txn 'get_all_updated_receipts' from sentinel context" warning ([\#3710](https://github.com/matrix-org/synapse/issues/3710))
- Fix bug where `state_cache` cache factor ignored environment variables ([\#3719](https://github.com/matrix-org/synapse/issues/3719))
- Fix bug in v0.33.3rc1 which caused infinite loops and OOMs ([\#3723](https://github.com/matrix-org/synapse/issues/3723))
- Fix bug introduced in v0.33.3rc1 which made the ToS give a 500 error ([\#3732](https://github.com/matrix-org/synapse/issues/3732))

Deprecations and Removals
-------------------------

- The Shared-Secret registration method of the legacy v1/register REST endpoint has been removed. For a replacement, please see [the admin/register API documentation](https://github.com/matrix-org/synapse/blob/master/docs/admin_api/register_api.rst). ([\#3703](https://github.com/matrix-org/synapse/issues/3703))

Internal Changes
----------------

- The test suite now can run under PostgreSQL. ([\#3423](https://github.com/matrix-org/synapse/issues/3423))
- Refactor HTTP replication endpoints to reduce code duplication ([\#3632](https://github.com/matrix-org/synapse/issues/3632))
- Tests now correctly execute on Python 3. ([\#3647](https://github.com/matrix-org/synapse/issues/3647))
- Sytests can now be run inside a Docker container. ([\#3660](https://github.com/matrix-org/synapse/issues/3660))
- Port over enough to Python 3 to allow the sytests to start. ([\#3668](https://github.com/matrix-org/synapse/issues/3668))
- Update docker base image from alpine 3.7 to 3.8. ([\#3669](https://github.com/matrix-org/synapse/issues/3669))
- Rename synapse.util.async to synapse.util.async_helpers to mitigate async becoming a keyword on Python 3.7. ([\#3678](https://github.com/matrix-org/synapse/issues/3678))
- Synapse's tests are now formatted with the black autoformatter. ([\#3679](https://github.com/matrix-org/synapse/issues/3679))
- Implemented a new testing base class to reduce test boilerplate. ([\#3684](https://github.com/matrix-org/synapse/issues/3684))
- Rename MAU prometheus metrics ([\#3690](https://github.com/matrix-org/synapse/issues/3690))
- add new error type ResourceLimit ([\#3707](https://github.com/matrix-org/synapse/issues/3707))
- Logcontexts for replication command handlers ([\#3709](https://github.com/matrix-org/synapse/issues/3709))
- Update admin register API documentation to reference a real user ID. ([\#3712](https://github.com/matrix-org/synapse/issues/3712))

1 files changed, 164 insertions, 47 deletions

diff --git a/synapse/http/site.py b/synapse/http/site.py
index 5fd30a4c2c..88ed3714f9 100644
--- a/synapse/http/site.py
+++ b/synapse/http/site.py
@@ -11,7 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 import contextlib
 import logging
 import time
@@ -19,8 +18,8 @@ import time
 from twisted.web.server import Request, Site
 
 from synapse.http import redact_uri
-from synapse.http.request_metrics import RequestMetrics
-from synapse.util.logcontext import ContextResourceUsage, LoggingContext
+from synapse.http.request_metrics import RequestMetrics, requests_counter
+from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
 
 logger = logging.getLogger(__name__)
 
@@ -34,25 +33,43 @@ class SynapseRequest(Request):
 
     It extends twisted's twisted.web.server.Request, and adds:
      * Unique request ID
+     * A log context associated with the request
      * Redaction of access_token query-params in __repr__
      * Logging at start and end
      * Metrics to record CPU, wallclock and DB time by endpoint.
 
-    It provides a method `processing` which should be called by the Resource
-    which is handling the request, and returns a context manager.
+    It also provides a method `processing`, which returns a context manager. If this
+    method is called, the request won't be logged until the context manager is closed;
+    this is useful for asynchronous request handlers which may go on processing the
+    request even after the client has disconnected.
 
+    Attributes:
+        logcontext(LoggingContext) : the log context for this request
     """
     def __init__(self, site, channel, *args, **kw):
         Request.__init__(self, channel, *args, **kw)
         self.site = site
-        self._channel = channel
+        self._channel = channel     # this is used by the tests
         self.authenticated_entity = None
         self.start_time = 0
 
+        # we can't yet create the logcontext, as we don't know the method.
+        self.logcontext = None
+
         global _next_request_seq
         self.request_seq = _next_request_seq
         _next_request_seq += 1
 
+        # whether an asynchronous request handler has called processing()
+        self._is_processing = False
+
+        # the time when the asynchronous request handler completed its processing
+        self._processing_finished_time = None
+
+        # what time we finished sending the response to the client (or the connection
+        # dropped)
+        self.finish_time = None
+
     def __repr__(self):
         # We overwrite this so that we don't log ``access_token``
         return '<%s at 0x%x method=%r uri=%r clientproto=%r site=%r>' % (
@@ -74,11 +91,116 @@ class SynapseRequest(Request):
         return self.requestHeaders.getRawHeaders(b"User-Agent", [None])[-1]
 
     def render(self, resrc):
+        # this is called once a Resource has been found to serve the request; in our
+        # case the Resource in question will normally be a JsonResource.
+
+        # create a LogContext for this request
+        request_id = self.get_request_id()
+        logcontext = self.logcontext = LoggingContext(request_id)
+        logcontext.request = request_id
+
         # override the Server header which is set by twisted
         self.setHeader("Server", self.site.server_version_string)
-        return Request.render(self, resrc)
+
+        with PreserveLoggingContext(self.logcontext):
+            # we start the request metrics timer here with an initial stab
+            # at the servlet name. For most requests that name will be
+            # JsonResource (or a subclass), and JsonResource._async_render
+            # will update it once it picks a servlet.
+            servlet_name = resrc.__class__.__name__
+            self._started_processing(servlet_name)
+
+            Request.render(self, resrc)
+
+            # record the arrival of the request *after*
+            # dispatching to the handler, so that the handler
+            # can update the servlet name in the request
+            # metrics
+            requests_counter.labels(self.method,
+                                    self.request_metrics.name).inc()
+
+    @contextlib.contextmanager
+    def processing(self):
+        """Record the fact that we are processing this request.
+
+        Returns a context manager; the correct way to use this is:
+
+        @defer.inlineCallbacks
+        def handle_request(request):
+            with request.processing("FooServlet"):
+                yield really_handle_the_request()
+
+        Once the context manager is closed, the completion of the request will be logged,
+        and the various metrics will be updated.
+        """
+        if self._is_processing:
+            raise RuntimeError("Request is already processing")
+        self._is_processing = True
+
+        try:
+            yield
+        except Exception:
+            # this should already have been caught, and sent back to the client as a 500.
+            logger.exception("Asynchronous messge handler raised an uncaught exception")
+        finally:
+            # the request handler has finished its work and either sent the whole response
+            # back, or handed over responsibility to a Producer.
+
+            self._processing_finished_time = time.time()
+            self._is_processing = False
+
+            # if we've already sent the response, log it now; otherwise, we wait for the
+            # response to be sent.
+            if self.finish_time is not None:
+                self._finished_processing()
+
+    def finish(self):
+        """Called when all response data has been written to this Request.
+
+        Overrides twisted.web.server.Request.finish to record the finish time and do
+        logging.
+        """
+        self.finish_time = time.time()
+        Request.finish(self)
+        if not self._is_processing:
+            with PreserveLoggingContext(self.logcontext):
+                self._finished_processing()
+
+    def connectionLost(self, reason):
+        """Called when the client connection is closed before the response is written.
+
+        Overrides twisted.web.server.Request.connectionLost to record the finish time and
+        do logging.
+        """
+        self.finish_time = time.time()
+        Request.connectionLost(self, reason)
+
+        # we only get here if the connection to the client drops before we send
+        # the response.
+        #
+        # It's useful to log it here so that we can get an idea of when
+        # the client disconnects.
+        with PreserveLoggingContext(self.logcontext):
+            logger.warn(
+                "Error processing request %r: %s %s", self, reason.type, reason.value,
+            )
+
+            if not self._is_processing:
+                self._finished_processing()
 
     def _started_processing(self, servlet_name):
+        """Record the fact that we are processing this request.
+
+        This will log the request's arrival. Once the request completes,
+        be sure to call finished_processing.
+
+        Args:
+            servlet_name (str): the name of the servlet which will be
+                processing this request. This is used in the metrics.
+
+                It is possible to update this afterwards by updating
+                self.request_metrics.name.
+        """
         self.start_time = time.time()
         self.request_metrics = RequestMetrics()
         self.request_metrics.start(
@@ -94,18 +216,32 @@ class SynapseRequest(Request):
         )
 
     def _finished_processing(self):
-        try:
-            context = LoggingContext.current_context()
-            usage = context.get_resource_usage()
-        except Exception:
-            usage = ContextResourceUsage()
+        """Log the completion of this request and update the metrics
+        """
+
+        if self.logcontext is None:
+            # this can happen if the connection closed before we read the
+            # headers (so render was never called). In that case we'll already
+            # have logged a warning, so just bail out.
+            return
+
+        usage = self.logcontext.get_resource_usage()
+
+        if self._processing_finished_time is None:
+            # we completed the request without anything calling processing()
+            self._processing_finished_time = time.time()
 
-        end_time = time.time()
+        # the time between receiving the request and the request handler finishing
+        processing_time = self._processing_finished_time - self.start_time
+
+        # the time between the request handler finishing and the response being sent
+        # to the client (nb may be negative)
+        response_send_time = self.finish_time - self._processing_finished_time
 
         # need to decode as it could be raw utf-8 bytes
         # from a IDN servname in an auth header
         authenticated_entity = self.authenticated_entity
-        if authenticated_entity is not None:
+        if authenticated_entity is not None and isinstance(authenticated_entity, bytes):
             authenticated_entity = authenticated_entity.decode("utf-8", "replace")
 
         # ...or could be raw utf-8 bytes in the User-Agent header.
@@ -116,22 +252,31 @@ class SynapseRequest(Request):
         user_agent = self.get_user_agent()
         if user_agent is not None:
             user_agent = user_agent.decode("utf-8", "replace")
+        else:
+            user_agent = "-"
+
+        code = str(self.code)
+        if not self.finished:
+            # we didn't send the full response before we gave up (presumably because
+            # the connection dropped)
+            code += "!"
 
         self.site.access_logger.info(
             "%s - %s - {%s}"
-            " Processed request: %.3fsec (%.3fsec, %.3fsec) (%.3fsec/%.3fsec/%d)"
+            " Processed request: %.3fsec/%.3fsec (%.3fsec, %.3fsec) (%.3fsec/%.3fsec/%d)"
             " %sB %s \"%s %s %s\" \"%s\" [%d dbevts]",
             self.getClientIP(),
             self.site.site_tag,
             authenticated_entity,
-            end_time - self.start_time,
+            processing_time,
+            response_send_time,
             usage.ru_utime,
             usage.ru_stime,
             usage.db_sched_duration_sec,
             usage.db_txn_duration_sec,
             int(usage.db_txn_count),
             self.sentLength,
-            self.code,
+            code,
             self.method,
             self.get_redacted_uri(),
             self.clientproto,
@@ -140,38 +285,10 @@ class SynapseRequest(Request):
         )
 
         try:
-            self.request_metrics.stop(end_time, self)
+            self.request_metrics.stop(self.finish_time, self)
         except Exception as e:
             logger.warn("Failed to stop metrics: %r", e)
 
-    @contextlib.contextmanager
-    def processing(self, servlet_name):
-        """Record the fact that we are processing this request.
-
-        Returns a context manager; the correct way to use this is:
-
-        @defer.inlineCallbacks
-        def handle_request(request):
-            with request.processing("FooServlet"):
-                yield really_handle_the_request()
-
-        This will log the request's arrival. Once the context manager is
-        closed, the completion of the request will be logged, and the various
-        metrics will be updated.
-
-        Args:
-            servlet_name (str): the name of the servlet which will be
-                processing this request. This is used in the metrics.
-
-                It is possible to update this afterwards by updating
-                self.request_metrics.servlet_name.
-        """
-        # TODO: we should probably just move this into render() and finish(),
-        # to save having to call a separate method.
-        self._started_processing(servlet_name)
-        yield
-        self._finished_processing()
-
 
 class XForwardedForRequest(SynapseRequest):
     def __init__(self, *args, **kw):
@@ -217,7 +334,7 @@ class SynapseSite(Site):
         proxied = config.get("x_forwarded", False)
         self.requestFactory = SynapseRequestFactory(self, proxied)
         self.access_logger = logging.getLogger(logger_name)
-        self.server_version_string = server_version_string
+        self.server_version_string = server_version_string.encode('ascii')
 
     def log(self, request):
         pass