#### Why I did it
Profiling the system state on init after fast-reboot during create_switch function execution, it is possible to see few python scripts running at the same time.
This parallel execution consume CPU time and the duration of create_switch is longer than it should be.
Following this finding, and the motivation to ensure these services will not interfere in the future, PMON is delayed in 90 seconds until the system finish the init flow after fastboot.
#### How I did it
Add a timer for PMON service.
Exclude for MLNX platform the start trigger of PMON when SYNCD starts in case of fastboot.
Copy the timer file to the host bin image.
#### How to verify it
Run fast-reboot on MLNX platform and observe faster create_switch execution time.
- Why I did it
Group all SONiC services together and able to manage them together. Will be used in config reload command as much simpler and generic way to restart services.
- How I did it
Add services to sonic.target
- How to verify it
Together with Azure/sonic-utilities#1199
config reload -y
Signed-off-by: Stepan Blyshchak <stepanb@nvidia.com>
When database service is down, psud daemon throws an error because of DB connection reset, this because pmon service has no dependency with database service.
To resolve this issue, added database service dependency to the pmon service.
Also, increased the net.core.somaxconn value to 512 to solve the connection failure on the scaled setup.
Issue Overview
shutdown flow
For any shutdown flow, which means all dockers are stopped in order, pmon docker stops after syncd docker has stopped, causing pmon docker fail to release sx_core resources and leaving sx_core in a bad state. The related logs are like the following:
INFO syncd.sh[23597]: modprobe: FATAL: Module sx_core is in use.
INFO syncd.sh[23597]: Unloading sx_core[FAILED]
INFO syncd.sh[23597]: rmmod: ERROR: Module sx_core is in use
config reload & service swss.restart
In the flows like "config reload" and "service swss restart", the failure cause further consequences:
sx_core initialization error with error message like "sx_core: create EMAD sdq 0 failed. err: -16"
syncd fails to execute the create switch api with error message "syncd_main: Runtime error: :- processEvent: failed to execute api: create, key: SAI_OBJECT_TYPE_SWITCH:oid:0x21000000000000, status: SAI_STATUS_FAILURE"
swss fails to call SAI API "SAI_SWITCH_ATTR_INIT_SWITCH", which causes orchagent to restart. This will introduce an extra 1 or 2 minutes for the system to be available, failing related test cases.
reboot, warm-reboot & fast-reboot
In the reboot flows including "reboot", "fast-reboot" and "warm-reboot" this failure doesn't have further negative effects since the system has already rebooted. In addition, "warm-reboot" requires the system to be shutdown as soon as possible to meet the GR time restriction of both BGP and LACP. "fast-reboot" also requires to meet the GR time restriction of BGP which is longer than LACP. In this sense, any unnecessary steps should be avoided. It's better to keep those flows untouched.
summary
To summarize, we have to come up with a way to ensure:
shutdown pmon docker ahead of syncd for "config reload" or "service swss restart" flow;
don't shutdown pmon docker ahead of syncd for "fast-reboot" or "warm-reboot" flow in order to save time.
for "reboot" flow, either order is acceptable.
Solution
To solve the issue, pmon shoud be stopped ahead of syncd stopped for all flows except for the warm-reboot.
- How I did it
To stop pmon ahead of syncd stopped. This is done in /usr/local/bin/syncd.sh::stop() and for all shutdown sequence.
Now pmon stops ahead of syncd so there must be a way in which pmon can start after syncd started. Another point that should be taken consideration is that pmon starting should be deferred so that services which have the logic of graceful restart in fast-reboot and warm-reboot have sufficient CPU cycles to meet their deadline.
This is done by add "syncd.service" as "After" to pmon.service and startin /usr/local/bin/syncd.sh::wait()
To start pmon automatically after syncd started.
- Move all minigraph-related action from rc.local to updategraph
- updategraph service is now after database. All feature services are now after and depending on updategraph
When Type=simple, systemd will consider the service activated immediately
after specified in ExecStart process is started. If there is downstream
service depending on the state prepared in ExecStart, there will be race condition.
For example, issue #390. In this case, database.service calls database.sh, which
calls docker run or docker start -a to start database container. However, systemd
considers database.service successfully started at the time database.sh begins,
not after docker run finishes. As database.service is consider started, bgp.service
can be started. The redis database, which bgp service depends on, might or might not
have been started at this time point.
To fix this issue (and still keeping the functionality to monitor docker status with
systemd), we split the ExecStart process into an ExecStartPre part and an ExecStart
part. docker run is splitted into docker run -d then docker attach , while docker start
-a is splitted into docker start and then docker attach. In this way, we make sure
the downstream services are blocked until container is successfully started.
