synapse-old/synapse/util/logcontext.py

579 lines
19 KiB
Python
Raw Normal View History

2016-01-06 21:26:29 -07:00
# Copyright 2014-2016 OpenMarket Ltd
2015-01-06 06:21:39 -07:00
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
2017-03-17 09:11:26 -06:00
""" Thread-local-alike tracking of log contexts within synapse
This module provides objects and utilities for tracking contexts through
synapse code, so that log lines can include a request identifier, and so that
CPU and database activity can be accounted for against the request that caused
them.
See doc/log_contexts.rst for details on how this works.
"""
2014-10-29 19:21:33 -06:00
import logging
2018-07-09 00:09:20 -06:00
import threading
from twisted.internet import defer
2014-10-29 19:21:33 -06:00
logger = logging.getLogger(__name__)
try:
import resource
# Python doesn't ship with a definition of RUSAGE_THREAD but it's defined
# to be 1 on linux so we hard code it.
RUSAGE_THREAD = 1
# If the system doesn't support RUSAGE_THREAD then this should throw an
# exception.
resource.getrusage(RUSAGE_THREAD)
2015-12-04 04:34:05 -07:00
def get_thread_resource_usage():
return resource.getrusage(RUSAGE_THREAD)
except Exception:
# If the system doesn't support resource.getrusage(RUSAGE_THREAD) then we
# won't track resource usage by returning None.
def get_thread_resource_usage():
return None
2014-10-30 04:13:46 -06:00
class ContextResourceUsage(object):
"""Object for tracking the resources used by a log context
Attributes:
ru_utime (float): user CPU time (in seconds)
ru_stime (float): system CPU time (in seconds)
db_txn_count (int): number of database transactions done
db_sched_duration_sec (float): amount of time spent waiting for a
database connection
db_txn_duration_sec (float): amount of time spent doing database
transactions (excluding scheduling time)
evt_db_fetch_count (int): number of events requested from the database
"""
__slots__ = [
"ru_stime", "ru_utime",
"db_txn_count", "db_txn_duration_sec", "db_sched_duration_sec",
"evt_db_fetch_count",
]
def __init__(self, copy_from=None):
"""Create a new ContextResourceUsage
Args:
copy_from (ContextResourceUsage|None): if not None, an object to
copy stats from
"""
if copy_from is None:
self.reset()
else:
self.ru_utime = copy_from.ru_utime
self.ru_stime = copy_from.ru_stime
self.db_txn_count = copy_from.db_txn_count
self.db_txn_duration_sec = copy_from.db_txn_duration_sec
self.db_sched_duration_sec = copy_from.db_sched_duration_sec
self.evt_db_fetch_count = copy_from.evt_db_fetch_count
def copy(self):
return ContextResourceUsage(copy_from=self)
def reset(self):
self.ru_stime = 0.
self.ru_utime = 0.
self.db_txn_count = 0
self.db_txn_duration_sec = 0
self.db_sched_duration_sec = 0
self.evt_db_fetch_count = 0
def __repr__(self):
return ("<ContextResourceUsage ru_stime='%r', ru_utime='%r', "
"db_txn_count='%r', db_txn_duration_sec='%r', "
"db_sched_duration_sec='%r', evt_db_fetch_count='%r'>") % (
self.ru_stime,
self.ru_utime,
self.db_txn_count,
self.db_txn_duration_sec,
self.db_sched_duration_sec,
self.evt_db_fetch_count,)
def __iadd__(self, other):
"""Add another ContextResourceUsage's stats to this one's.
Args:
other (ContextResourceUsage): the other resource usage object
"""
self.ru_utime += other.ru_utime
self.ru_stime += other.ru_stime
self.db_txn_count += other.db_txn_count
self.db_txn_duration_sec += other.db_txn_duration_sec
self.db_sched_duration_sec += other.db_sched_duration_sec
self.evt_db_fetch_count += other.evt_db_fetch_count
return self
def __isub__(self, other):
self.ru_utime -= other.ru_utime
self.ru_stime -= other.ru_stime
self.db_txn_count -= other.db_txn_count
self.db_txn_duration_sec -= other.db_txn_duration_sec
self.db_sched_duration_sec -= other.db_sched_duration_sec
self.evt_db_fetch_count -= other.evt_db_fetch_count
return self
def __add__(self, other):
res = ContextResourceUsage(copy_from=self)
res += other
return res
def __sub__(self, other):
res = ContextResourceUsage(copy_from=self)
res -= other
return res
2014-10-29 19:21:33 -06:00
class LoggingContext(object):
2014-10-30 04:13:46 -06:00
"""Additional context for log formatting. Contexts are scoped within a
2016-02-10 04:29:21 -07:00
"with" block.
If a parent is given when creating a new context, then:
- logging fields are copied from the parent to the new context on entry
- when the new context exits, the cpu usage stats are copied from the
child to the parent
2014-10-30 04:13:46 -06:00
Args:
name (str): Name for the context for debugging.
parent_context (LoggingContext|None): The parent of the new context
2014-10-30 04:13:46 -06:00
"""
__slots__ = [
"previous_context", "name", "parent_context",
"_resource_usage",
2018-05-22 04:16:07 -06:00
"usage_start",
"main_thread", "alive",
"request", "tag",
]
2014-10-29 19:21:33 -06:00
thread_local = threading.local()
class Sentinel(object):
2014-10-30 04:13:46 -06:00
"""Sentinel to represent the root context"""
2014-10-29 19:21:33 -06:00
__slots__ = []
2014-10-30 04:13:46 -06:00
def __str__(self):
return "sentinel"
2014-10-29 19:21:33 -06:00
def copy_to(self, record):
pass
def start(self):
pass
def stop(self):
pass
def add_database_transaction(self, duration_sec):
pass
def add_database_scheduled(self, sched_sec):
pass
def record_event_fetch(self, event_count):
pass
2016-02-03 06:51:25 -07:00
def __nonzero__(self):
return False
__bool__ = __nonzero__ # python3
2016-02-03 06:51:25 -07:00
2014-10-29 19:21:33 -06:00
sentinel = Sentinel()
def __init__(self, name=None, parent_context=None):
2016-02-10 04:25:19 -07:00
self.previous_context = LoggingContext.current_context()
2014-10-29 19:21:33 -06:00
self.name = name
# track the resources used by this context so far
self._resource_usage = ContextResourceUsage()
2018-05-22 04:16:07 -06:00
# If alive has the thread resource usage when the logcontext last
# became active.
self.usage_start = None
2018-05-22 04:16:07 -06:00
self.main_thread = threading.current_thread()
self.request = None
2016-02-03 06:51:25 -07:00
self.tag = ""
2016-02-04 03:22:44 -07:00
self.alive = True
2014-10-29 19:21:33 -06:00
self.parent_context = parent_context
2014-10-29 19:21:33 -06:00
def __str__(self):
2014-10-30 04:13:46 -06:00
return "%s@%x" % (self.name, id(self))
2014-10-29 19:21:33 -06:00
@classmethod
def current_context(cls):
"""Get the current logging context from thread local storage
Returns:
LoggingContext: the current logging context
"""
2014-10-29 19:21:33 -06:00
return getattr(cls.thread_local, "current_context", cls.sentinel)
@classmethod
def set_current_context(cls, context):
"""Set the current logging context in thread local storage
Args:
context(LoggingContext): The context to activate.
Returns:
The context that was previously active
"""
current = cls.current_context()
2016-02-04 03:22:44 -07:00
if current is not context:
current.stop()
cls.thread_local.current_context = context
context.start()
return current
2014-10-29 19:21:33 -06:00
def __enter__(self):
2014-10-30 04:13:46 -06:00
"""Enters this logging context into thread local storage"""
2016-02-10 04:23:32 -07:00
old_context = self.set_current_context(self)
2016-02-10 04:25:19 -07:00
if self.previous_context != old_context:
2016-02-10 04:23:32 -07:00
logger.warn(
2016-02-10 04:25:19 -07:00
"Expected previous context %r, found %r",
self.previous_context, old_context
2016-02-10 04:23:32 -07:00
)
2016-02-04 03:22:44 -07:00
self.alive = True
if self.parent_context is not None:
self.parent_context.copy_to(self)
2014-10-29 19:21:33 -06:00
return self
def __exit__(self, type, value, traceback):
2014-10-30 04:13:46 -06:00
"""Restore the logging context in thread local storage to the state it
was before this context was entered.
Returns:
2018-05-22 04:16:07 -06:00
None to avoid suppressing any exceptions that were thrown.
2014-10-30 04:13:46 -06:00
"""
2016-02-10 04:25:19 -07:00
current = self.set_current_context(self.previous_context)
if current is not self:
if current is self.sentinel:
logger.warn("Expected logging context %s has been lost", self)
else:
logger.warn(
"Current logging context %s is not expected context %s",
current,
self
)
2016-02-10 04:25:19 -07:00
self.previous_context = None
2016-02-04 03:22:44 -07:00
self.alive = False
2014-10-29 19:21:33 -06:00
# if we have a parent, pass our CPU usage stats on
if self.parent_context is not None:
self.parent_context._resource_usage += self._resource_usage
# reset them in case we get entered again
self._resource_usage.reset()
2014-10-29 19:21:33 -06:00
def copy_to(self, record):
"""Copy logging fields from this context to a log record or
another LoggingContext
"""
2014-10-29 19:21:33 -06:00
# 'request' is the only field we currently use in the logger, so that's
# all we need to copy
record.request = self.request
def start(self):
if threading.current_thread() is not self.main_thread:
2018-05-22 04:16:07 -06:00
logger.warning("Started logcontext %s on different thread", self)
return
2018-05-22 04:16:07 -06:00
# If we haven't already started record the thread resource usage so
# far
if not self.usage_start:
self.usage_start = get_thread_resource_usage()
def stop(self):
if threading.current_thread() is not self.main_thread:
2018-05-22 04:16:07 -06:00
logger.warning("Stopped logcontext %s on different thread", self)
return
# When we stop, let's record the cpu used since we started
if not self.usage_start:
logger.warning(
"Called stop on logcontext %s without calling start", self,
)
return
2018-05-22 04:16:07 -06:00
usage_end = get_thread_resource_usage()
2018-05-22 04:16:07 -06:00
self._resource_usage.ru_utime += usage_end.ru_utime - self.usage_start.ru_utime
self._resource_usage.ru_stime += usage_end.ru_stime - self.usage_start.ru_stime
self.usage_start = None
def get_resource_usage(self):
"""Get resources used by this logcontext so far.
2018-05-22 04:16:07 -06:00
Returns:
ContextResourceUsage: a *copy* of the object tracking resource
usage so far
2018-05-22 04:16:07 -06:00
"""
# we always return a copy, for consistency
res = self._resource_usage.copy()
2018-05-22 04:16:07 -06:00
# If we are on the correct thread and we're currently running then we
# can include resource usage so far.
is_main_thread = threading.current_thread() is self.main_thread
if self.alive and self.usage_start and is_main_thread:
current = get_thread_resource_usage()
res.ru_utime += current.ru_utime - self.usage_start.ru_utime
res.ru_stime += current.ru_stime - self.usage_start.ru_stime
return res
def add_database_transaction(self, duration_sec):
self._resource_usage.db_txn_count += 1
self._resource_usage.db_txn_duration_sec += duration_sec
def add_database_scheduled(self, sched_sec):
"""Record a use of the database pool
Args:
sched_sec (float): number of seconds it took us to get a
connection
"""
self._resource_usage.db_sched_duration_sec += sched_sec
def record_event_fetch(self, event_count):
"""Record a number of events being fetched from the db
Args:
event_count (int): number of events being fetched
"""
self._resource_usage.evt_db_fetch_count += event_count
2014-10-29 19:21:33 -06:00
class LoggingContextFilter(logging.Filter):
2014-10-30 04:13:46 -06:00
"""Logging filter that adds values from the current logging context to each
record.
Args:
**defaults: Default values to avoid formatters complaining about
missing fields
"""
2014-10-29 19:21:33 -06:00
def __init__(self, **defaults):
self.defaults = defaults
def filter(self, record):
2014-10-30 04:13:46 -06:00
"""Add each fields from the logging contexts to the record.
Returns:
True to include the record in the log output.
"""
2014-10-29 19:21:33 -06:00
context = LoggingContext.current_context()
for key, value in self.defaults.items():
setattr(record, key, value)
context.copy_to(record)
return True
class PreserveLoggingContext(object):
2014-10-30 04:13:46 -06:00
"""Captures the current logging context and restores it when the scope is
exited. Used to restore the context after a function using
@defer.inlineCallbacks is resumed by a callback from the reactor."""
2016-02-04 03:22:44 -07:00
__slots__ = ["current_context", "new_context", "has_parent"]
def __init__(self, new_context=LoggingContext.sentinel):
self.new_context = new_context
2014-10-30 04:13:46 -06:00
2014-10-29 19:21:33 -06:00
def __enter__(self):
2014-10-30 04:13:46 -06:00
"""Captures the current logging context"""
self.current_context = LoggingContext.set_current_context(
self.new_context
)
2014-10-29 19:21:33 -06:00
2016-02-04 03:22:44 -07:00
if self.current_context:
2016-02-10 04:25:19 -07:00
self.has_parent = self.current_context.previous_context is not None
2016-02-04 03:22:44 -07:00
if not self.current_context.alive:
2016-02-09 02:20:06 -07:00
logger.debug(
2016-02-04 03:22:44 -07:00
"Entering dead context: %s",
self.current_context,
)
2014-10-29 19:21:33 -06:00
def __exit__(self, type, value, traceback):
2014-10-30 04:13:46 -06:00
"""Restores the current logging context"""
2016-02-04 03:22:44 -07:00
context = LoggingContext.set_current_context(self.current_context)
if context != self.new_context:
logger.warn(
2016-02-04 03:22:44 -07:00
"Unexpected logging context: %s is not %s",
context, self.new_context,
)
2015-05-08 12:53:34 -06:00
if self.current_context is not LoggingContext.sentinel:
2016-02-04 03:22:44 -07:00
if not self.current_context.alive:
2016-02-09 02:20:06 -07:00
logger.debug(
2015-05-08 12:53:34 -06:00
"Restoring dead context: %s",
self.current_context,
)
def preserve_fn(f):
"""Function decorator which wraps the function with run_in_background"""
def g(*args, **kwargs):
return run_in_background(f, *args, **kwargs)
return g
def run_in_background(f, *args, **kwargs):
"""Calls a function, ensuring that the current context is restored after
return from the function, and that the sentinel context is set once the
deferred returned by the function completes.
Useful for wrapping functions that return a deferred which you don't yield
on (for instance because you want to pass it to deferred.gatherResults()).
Note that if you completely discard the result, you should make sure that
`f` doesn't raise any deferred exceptions, otherwise a scary-looking
CRITICAL error about an unhandled error will be logged without much
indication about where it came from.
"""
current = LoggingContext.current_context()
try:
res = f(*args, **kwargs)
except: # noqa: E722
# the assumption here is that the caller doesn't want to be disturbed
# by synchronous exceptions, so let's turn them into Failures.
return defer.fail()
if not isinstance(res, defer.Deferred):
return res
if res.called and not res.paused:
# The function should have maintained the logcontext, so we can
# optimise out the messing about
return res
# The function may have reset the context before returning, so
# we need to restore it now.
ctx = LoggingContext.set_current_context(current)
# The original context will be restored when the deferred
# completes, but there is nothing waiting for it, so it will
# get leaked into the reactor or some other function which
# wasn't expecting it. We therefore need to reset the context
# here.
#
# (If this feels asymmetric, consider it this way: we are
# effectively forking a new thread of execution. We are
# probably currently within a ``with LoggingContext()`` block,
# which is supposed to have a single entry and exit point. But
# by spawning off another deferred, we are effectively
# adding a new exit point.)
res.addBoth(_set_context_cb, ctx)
return res
2016-02-04 03:22:44 -07:00
def make_deferred_yieldable(deferred):
"""Given a deferred, make it follow the Synapse logcontext rules:
If the deferred has completed (or is not actually a Deferred), essentially
does nothing (just returns another completed deferred with the
result/failure).
If the deferred has not yet completed, resets the logcontext before
returning a deferred. Then, when the deferred completes, restores the
current logcontext before running callbacks/errbacks.
(This is more-or-less the opposite operation to run_in_background.)
"""
if not isinstance(deferred, defer.Deferred):
return deferred
if deferred.called and not deferred.paused:
# it looks like this deferred is ready to run any callbacks we give it
# immediately. We may as well optimise out the logcontext faffery.
return deferred
# ok, we can't be sure that a yield won't block, so let's reset the
# logcontext, and add a callback to the deferred to restore it.
prev_context = LoggingContext.set_current_context(LoggingContext.sentinel)
deferred.addBoth(_set_context_cb, prev_context)
return deferred
def _set_context_cb(result, context):
"""A callback function which just sets the logging context"""
LoggingContext.set_current_context(context)
return result
2016-02-04 03:22:44 -07:00
# modules to ignore in `logcontext_tracer`
_to_ignore = [
"synapse.util.logcontext",
"synapse.http.server",
"synapse.storage._base",
"synapse.util.async",
]
def logcontext_tracer(frame, event, arg):
"""A tracer that logs whenever a logcontext "unexpectedly" changes within
a function. Probably inaccurate.
Use by calling `sys.settrace(logcontext_tracer)` in the main thread.
"""
if event == 'call':
name = frame.f_globals["__name__"]
if name.startswith("synapse"):
if name == "synapse.util.logcontext":
if frame.f_code.co_name in ["__enter__", "__exit__"]:
tracer = frame.f_back.f_trace
if tracer:
tracer.just_changed = True
tracer = frame.f_trace
if tracer:
return tracer
if not any(name.startswith(ig) for ig in _to_ignore):
return LineTracer()
class LineTracer(object):
__slots__ = ["context", "just_changed"]
def __init__(self):
self.context = LoggingContext.current_context()
self.just_changed = False
def __call__(self, frame, event, arg):
if event in 'line':
if self.just_changed:
self.context = LoggingContext.current_context()
self.just_changed = False
else:
c = LoggingContext.current_context()
if c != self.context:
logger.info(
"Context changed! %s -> %s, %s, %s",
self.context, c,
frame.f_code.co_filename, frame.f_lineno
)
self.context = c
return self