[1.x] system under heavy load stops to handle calls wia ws api

[spscream]

What version of Janus is this happening on?
master

Have you tested a more recent version of Janus too?
we use master

Was this working before?
we had no issue on lower versions, but we had another issue with memory leaks on them, so we got OOM some time.

Is there a gdb or libasan trace of the issue?
no

Under some circuumstances janus stops to handle any api calls via ws - ws is accepting connections, but create_session has timeouts and janus doesn't work. We tried to add debug_locks, but we got situation what janus server eats all memory and IO and stuck, so we got problems with performance before issue appeared. This happens on one of our customers with very heavy load on januses.

How can we debug this and what can be a cause of it?

https://gist.github.com/spscream/84aa7bca6f8e3f43e07d4c58f414e9cd - recent log of such situation with debug_level = 4

[spscream]

may we have some partial debug_locks implementation only for locks on session creation to debug it?

[atoppi]

Under some circuumstances janus stops to handle any api calls via ws

Do other requests work though? E.g. "info" request or attaching a handle for a different plugin. That could help understanding if the deadlock lies in the transport (WebSocket) or somewhere else.

but create_session has timeouts and janus doesn't work

Do you mean plugin->create_session? If that's the case and you are using the videoroom plugin, then it's very likely a deadlock on videoroom sessions_mutex.

We tried to add debug_locks, but we got situation what janus server eats all memory and IO and stuck, so we got problems with performance before issue appeared

The debug_locks has a massive impact on verbosity, that's probably the reason of performance issue. If you are using Janus in a containerized environment with cgroups v2, the huge log file increasing might increase the memory allocated (due to pages being kept in buffer) and might explain the OOM.

may we have some partial debug_locks implementation only for locks on session creation to debug it?

There is no such option available. You might try customizing the code and just logs the sessions_mutex lock/unlock (given that my previous guess is correct)

[spscream]

Do you mean plugin->create_session? If that's the case and you are using the videoroom plugin, then it's very likely a deadlock on videoroom sessions_mutex.

I mean the following request isn't working

{
        "janus" : "create",
        "transaction" : "<random alphanumeric string>"
}

Do other requests work though? E.g. "info" request or attaching a handle for a different plugin. That could help understanding if the deadlock lies in the transport (WebSocket) or somewhere else.

I didn't tried this, but I can try next time this happens.

[atoppi]

Taking a deeper inspection at the logs you shared, the issue seems to start with some errors:

Didn't receive audio for more than 1 second(s)...
Didn't receive video for more than 1 second(s)...

SRTCP unprotect error: srtp_err_status_replay_fail

Error setting ICE locally

Those reminds me of situations where the host memory is exhausting (like the issue about cgroups v2 I already mentioned).

Are you running janus in containers with a memory limit?
Are you doing long lasting recordings?

If you suspect a memory leak, try running your janus app in a lab environment under valgrind.

[spscream]

Are you running janus in containers with a memory limit?

we are running it in containers, but without memory limits set

Are you doing long lasting recordings?

yes

[atoppi]

All right, this is a long shot, but can you check the status of the memory in the containers?

cat /sys/fs/cgroup/system.slice/docker-<long ID>.scope/memory.max
cat /sys/fs/cgroup/system.slice/docker-<long ID>.scope/memory.stat

replace long-id with the id of the docker container.

If you see the file bytes amount approaching the memory.max limit, then containers under cgroups v2 will start having issues with memory allocation, since they share the same memory for network and file buffers.
We have proposed a potential patch (in the PR) for long lasting recordings where basically we flush the buffers and tells the kernel to release the used pages every ~2MB of written data.

[spscream]

Got this on another customer. We are using this on our production, so in case we got it on our servers, we will enable debug_locks, since we have no such high load as on customers servers, we will try to catch it.

[spscream]

got info from our customer, after janus restart

# cat /sys/fs/cgroup/system.slice/docker-9efbad4747099c922f9d8c1dc16a74e029e871a1e21cdcea1b5b8bfc7de47546.scope/memory.max
max
# cat /sys/fs/cgroup/system.slice/docker-9efbad4747099c922f9d8c1dc16a74e029e871a1e21cdcea1b5b8bfc7de47546.scope/memory.stat
anon 21372928
file 4861952
kernel_stack 294912
pagetables 131072
percpu 288
sock 0
shmem 0
file_mapped 0
file_dirty 0
file_writeback 0
swapcached 0
anon_thp 0
file_thp 0
shmem_thp 0
inactive_anon 21368832
active_anon 4096
inactive_file 4861952
active_file 0
unevictable 0
slab_reclaimable 444944
slab_unreclaimable 357096
slab 802040
workingset_refault_anon 0
workingset_refault_file 0
workingset_activate_anon 0
workingset_activate_file 0
workingset_restore_anon 0
workingset_restore_file 0
workingset_nodereclaim 0
pgfault 48048
pgmajfault 0
pgrefill 0
pgscan 0
pgsteal 0
pgactivate 0
pgdeactivate 0
pglazyfree 0
pglazyfreed 0
thp_fault_alloc 0
thp_collapse_alloc 0

[spscream]

We will try to reproduce it on our lab with enabling recording and setting memory limits on docker container.

[spscream]

we unable to reproduce it on our test stand, but it consistently repeats on customers envs

[atoppi]

got info from our customer, after janus restart

Those data are useless after a restart, we need them while the issue exists (btw, the max output is wrong, it should be a number).

we unable to reproduce it on our test stand, but it consistently repeats on customers envs

If you suspect a deadlock wait for the issue and then provide the output of this gdb snippet:

gdb -p "$(pidof janus)" --batch -ex "set print pretty on" -ex "set pagination off" -ex "thread apply all bt full"