On 8/26/2024 11:10 AM, Manoj Vishwanathan wrote:

From: Marco Leogrande <redacted>

When a workqueue is created with `WQ_UNBOUND`, its work items are
served by special worker-pools, whose host workers are not bound to
any specific CPU. In the default configuration (i.e. when
`queue_delayed_work` and friends do not specify which CPU to run the
work item on), `WQ_UNBOUND` allows the work item to be executed on any
CPU in the same node of the CPU it was enqueued on. While this
solution potentially sacrifices locality, it avoids contention with
other processes that might dominate the CPU time of the processor the
work item was scheduled on.

This is not just a theoretical problem: in a praticular scenario

Nit: s/praticular/particular/

misconfigured process was hogging most of the time from CPU0, leaving
less than 0.5% of its CPU time to the kworker. The IDPF workqueues
that were using the kworker on CPU0 suffered large completion delays
as a result, causing performance degradation, timeouts and eventual
system crash.

Curious how the delay could result in a full system crash. That seems
like some other concurrency issue. I guess something like a Tx timeout
could happen though.

Tested:

* I have also run a manual test to gauge the performance
  improvement. The test consists of an antagonist process
  (`./stress --cpu 2`) consuming as much of CPU 0 as possible. This
  process is run under `taskset 01` to bind it to CPU0, and its
  priority is changed with `chrt -pQ 9900 10000 ${pid}` and
  `renice -n -20 ${pid}` after start.

  Then, the IDPF driver is forced to prefer CPU0 by editing all calls
  to `queue_delayed_work`, `mod_delayed_work`, etc... to use CPU 0.

  Finally, `ktraces` for the workqueue events are collected.

  Without the current patch, the antagonist process can force
  arbitrary delays between `workqueue_queue_work` and
  `workqueue_execute_start`, that in my tests were as high as
  `30ms`. With the current patch applied, the workqueue can be
  migrated to another unloaded CPU in the same node, and, keeping
  everything else equal, the maximum delay I could see was `6us`.

Hmm. I don't have a direct issue with using WQ_UNBOUND, and I can't
think of any reason these work queue tasks *need* to be CPU bound.

I do feel like there may be other solutions to managing the tasks on the
system such that this isn't necessary.

However, if using WQ_UNBOUND solves these problems and is simpler in
that system administrators are less likely to screw things up, I think
its a net positive.

I do not know if there are any other side effects of WQ_UNBOUND, so take
this with a grain of salt:
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>

This seems like quite a lot of work queues for a driver :D