Re: [PATCH v2 0/5] Introduce QPW for per-cpu operations (v2)

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Mar 05, 2026 at 05:55:12PM +0100, Frederic Weisbecker wrote:
> Le Mon, Mar 02, 2026 at 12:49:45PM -0300, Marcelo Tosatti a écrit :
> > The problem:
> > Some places in the kernel implement a parallel programming strategy
> > consisting on local_locks() for most of the work, and some rare remote
> > operations are scheduled on target cpu. This keeps cache bouncing low since
> > cacheline tends to be mostly local, and avoids the cost of locks in non-RT
> > kernels, even though the very few remote operations will be expensive due
> > to scheduling overhead.
> > 
> > On the other hand, for RT workloads this can represent a problem: getting
> > an important workload scheduled out to deal with remote requests is
> > sure to introduce unexpected deadline misses.
> > 
> > The idea:
> > Currently with PREEMPT_RT=y, local_locks() become per-cpu spinlocks.
> > In this case, instead of scheduling work on a remote cpu, it should
> > be safe to grab that remote cpu's per-cpu spinlock and run the required
> > work locally. That major cost, which is un/locking in every local function,
> > already happens in PREEMPT_RT.
> > 
> > Also, there is no need to worry about extra cache bouncing:
> > The cacheline invalidation already happens due to schedule_work_on().
> > 
> > This will avoid schedule_work_on(), and thus avoid scheduling-out an
> > RT workload.
> > 
> > Proposed solution:
> > A new interface called Queue PerCPU Work (QPW), which should replace
> > Work Queue in the above mentioned use case.
> > 
> > If CONFIG_QPW=n this interfaces just wraps the current
> > local_locks + WorkQueue behavior, so no expected change in runtime.
> > 
> > If CONFIG_QPW=y, and qpw kernel boot option =1, 
> > queue_percpu_work_on(cpu,...) will lock that cpu's per-cpu structure
> > and perform work on it locally. This is possible because on 
> > functions that can be used for performing remote work on remote 
> > per-cpu structures, the local_lock (which is already
> > a this_cpu spinlock()), will be replaced by a qpw_spinlock(), which
> > is able to get the per_cpu spinlock() for the cpu passed as parameter.
> 
> So let me summarize what are the possible design solutions, on top of our discussions,
> so we can compare:

I find this summary difficult to comprehend. The way i see it is:

A certain class of data structures can be manipulated only by each individual CPU (the
per-CPU caches), since they lack proper locks for such data to be
manipulated by remote CPUs.

There are certain operations which require such data to be manipulated,
therefore work is queued to execute on the owner CPUs.

> 
> 1) Never queue remotely but always queue locally and execute on userspace

When you say "queue locally", do you mean to queue the data structure 
manipulation to happen on return to userspace of the owner CPU ?

What if it does not return to userspace ? (or takes a long time to return 
to userspace?).

>    return via task work.
> 
>    Pros:
>          - Simple and easy to maintain.
> 
>    Cons:
>          - Need a case by case handling.
> 
> 	 - Might be suitable for full userspace applications but not for
>            some HPC usecases. In the best world MPI is fully implemented in
>            userspace but that doesn't appear to be the case.
> 
> 2) Queue locally the workqueue right away or do it remotely (if it's
>    really necessary) if the isolated CPU is in userspace, otherwise queue
>    it for execution on return to kernel. The work will be handled by preemption
>    to a worker or by a workqueue flush on return to userspace.
> 
>    Pros:
>         - The local queue handling is simple.
> 
>    Cons:
>         - The remote queue must synchronize with return to userspace and
> 	  eventually postpone to return to kernel if the target is in userspace.
> 	  Also it may need to differentiate IRQs and syscalls.
> 
>         - Therefore still involve some case by case handling eventually.
>    
>         - Flushing the global workqueues to avoid deadlocks is unadvised as shown
>           in the comment above flush_scheduled_work(). It even triggers a
>           warning. Significant efforts have been put to convert all the existing
> 	  users. It's not impossible to sell in our case because we shouldn't
> 	  hold a lock upon return to userspace. But that will restore a new
> 	  dangerous API.
> 
>         - Queueing the workqueue / flushing involves a context switch which
>           induce more noise (eg: tick restart)
> 	  
>         - As above, probably not suitable for HPC.
> 
> 3) QPW: Handle the work remotely
> 
>    Pros:
>         - Works on all cases, without any surprise.
> 
>    Cons:
>         - Introduce new locking scheme to maintain and debug.
> 
>         - Needs case by case handling.
> 
> Thoughts?
> 
> -- 
> Frederic Weisbecker
> SUSE Labs

Its hard for me to parse your concise summary (perhaps it could be more
verbose).

Anyway, one thought is to use some sort of SRCU type protection on the 
per-CPU caches.
But that adds cost as well (compared to non-SRCU), which then seems to
have cost similar to adding per-CPU spinlocks.