________________________________________
From: Christian Brauner <brauner@kernel.org>
Sent: Thursday, March 9, 2023 9:05 AM
To: Anjali Kulkarni
Cc: davem@davemloft.net; edumazet@google.com; kuba@kernel.org; pabeni@redhat.com; zbr@ioremap.net; johannes@sipsolutions.net; ecree.xilinx@gmail.com; leon@kernel.org; keescook@chromium.org; socketcan@hartkopp.net; petrm@nvidia.com; linux-kernel@vger.kernel.org; netdev@vger.kernel.org
Subject: Re: [PATCH 0/5] Process connector bug fixes & enhancements

On Wed, Mar 08, 2023 at 07:19:48PM -0800, Anjali Kulkarni wrote:

From: Anjali Kulkarni <redacted>

In this series, we add back filtering to the proc connector module. This
is required to fix some bugs and also will enable the addition of event
based filtering, which will improve performance for anyone interested
in a subset of process events, as compared to the current approach,
which is to send all event notifications.

Thus, a client can register to listen for only exit or fork or a mix or
all of the events. This greatly enhances performance - currently, we
need to listen to all events, and there are 9 different types of events.
For eg. handling 3 types of events - 8K-forks + 8K-exits + 8K-execs takes
200ms, whereas handling 2 types - 8K-forks + 8K-exits takes about 150ms,
and handling just one type - 8K exits takes about 70ms.

Reason why we need the above changes and also a new event type
PROC_EVENT_NONZERO_EXIT, which is only sent by kernel to a listening
application when any process exiting has a non-zero exit status is:

Oracle DB runs on a large scale with 100000s of short lived processes,
starting up and exiting quickly. A process monitoring DB daemon which
tracks and cleans up after processes that have died without a proper exit
needs notifications only when a process died with a non-zero exit code
(which should be rare).

This change will give Oracle DB substantial performance savings - it takes
50ms to scan about 8K PIDs in /proc, about 500ms for 100K PIDs. DB does
this check every 3 secs, so over an hour we save 10secs for 100K PIDs.

Measuring the time using pidfds for monitoring 8K process exits took 4
times longer - 200ms, as compared to 70ms using only exit notifications
of proc connector. Hence, we cannot use pidfd for our use case.

Just out of curiosity, what's the reason this took so much longer?

ANJALI> I have not looked in it in detail, but it seems this may be due to the number of system calls involved. The monitored process needs to send it’s pidfd to the monitoring process, which adds the pidfd in an epoll interface and removes it on process exit. (I did not include time required from monitored process’s side, to open the pidfd and send it, in this). For our case, we cannot have our monitoring process know about every exit (or receive new process’s fd) that happens due to the large no. of exits happening.

This kind of a new event could also be useful to other applications like
Google's lmkd daemon, which needs a killed process's exit notification.

Fwiw - independent of this thing here - I think we might need to also
think about making the exit status of a process readable from a pidfd.
Even after the process has been exited + reaped... I have a _rough_ idea
how I thought this could work:

* introduce struct pidfd_info
* allocate one struct pidfd_info per struct pid _lazily_when the first a pidfd is created
* stash struct pidfd_info in pidfd_file->private_data
* add .exit_status field to struct pidfd_info
* when process exits statsh exit status in struct pidfd_info
* add either new system call or ioctl() to pidfd which returns EAGAIN or
  sm until process has exited and then becomes readable

Thought needs to be put into finding struct pidfd_info based on struct pid...

ANJALI> This seems like a useful feature to have.