Fix: consumerd: slow metadata push slows down application registration
Issue observed
--------------
When rotating the channels of a session configured with a "per-pid"
buffer sharing policy, applications with a long registration
timeout (e.g. LTTNG_UST_REGISTER_TIMEOUT=-1, see LTTNG-UST(3)) sometimes
experience long start-up times.
Cause
-----
The session list lock is held during the registration of an application
and during the setup of a rotation.
While setting up a rotation in the userspace domain, the session daemon
flushes its metadata cache to the userspace consumer daemon and waits
for a confirmation that all metadata emitted before that point in time
has been serialized (whether on disk or sent through a network output).
As the consumer daemon waits for the metadata to be consumed, it
periodically checks the metadata stream's output position with a 200ms
delay (see DEFAULT_METADATA_AVAILABILITY_WAIT_TIME).
In practice, in per-uid mode, this delay is seldomly encountered since
the metadata has already been pushed by the consumption thread.
Moreover, if it was not, a single polling iteration will typically
suffice.
However, in per-pid buffering mode and with a sustained "heavy" data
production rate, this delay becomes problematic since:
- metadata is pushed for every application,
- the delay is hit almost systematically as the consumption thread
is busy and has to catch up to consume the most recent metadata.
Hence, some rotation setups can easily take multiple seconds (at least
200ms per application). This makes the locking scheme employed on that
path unsuitable as it blocks some operations (like application
registrations) for an extended period of time.
Solution
--------
The polling "back-off" delay is eliminated by using a waiter that allows
the consumer daemon thread that runs the metadata push command to
wake-up whenever the criteria used to evaluate the "pushed" metadata
position are changed.
Those criteria are:
- the metadata stream's pushed position
- the lifetime of the metadata channel's stream
- the status of the session's endpoint
Whenever those states are affected, the waiters are woken-up to force a
re-evaluation of the metadata cache flush position and, eventually,
cause the metadata push command to complete.
Note
----
The waiter queue is adapted from urcu-wait.h of liburcu (also LGPL
licensed).
Change-Id: Ib86c2e878abe205c73f930e6de958c0b10486a37
Signed-off-by: Jérémie Galarneau <jeremie.galarneau@efficios.com>
This page took 0.026812 seconds and 4 git commands to generate.