Re: [PATCH v2] rcu: Reduce synchronize_rcu() latency by reporting GP kthread's CPU QS early

[Joel Fernandes <joelagnelf@nvidia.com>]

On Sun, Dec 28, 2025 at 08:37:14PM -0800, Paul E. McKenney wrote:
> On Sun, Dec 28, 2025 at 09:49:45PM -0500, Joel Fernandes wrote:
> > 
> > 
> > > On Dec 28, 2025, at 7:04 PM, Paul E. McKenney <paulmck@kernel.org> wrote:
> > > 
> > > On Sun, Dec 28, 2025 at 06:57:58PM +0100, Uladzislau Rezki wrote:
> > >>> On Thu, Dec 25, 2025 at 09:33:39PM -0500, Joel Fernandes wrote:
> > >>> On Thu, Dec 25, 2025 at 10:35:44AM -0800, Paul E. McKenney wrote:
> > >>>> On Mon, Dec 22, 2025 at 10:46:29PM -0500, Joel Fernandes wrote:
> > >>>>> The RCU grace period mechanism uses a two-phase FQS (Force Quiescent
> > >>>>> State) design where the first FQS saves dyntick-idle snapshots and
> > >>>>> the second FQS compares them. This results in long and unnecessary latency
> > >>>>> for synchronize_rcu() on idle systems (two FQS waits of ~3ms each with
> > >>>>> 1000HZ) whenever one FQS wait sufficed.
> > >>>>> 
> > >>>>> Some investigations showed that the GP kthread's CPU is the holdout CPU
> > >>>>> a lot of times after the first FQS as - it cannot be detected as "idle"
> > >>>>> because it's actively running the FQS scan in the GP kthread.
> > >>>>> 
> > >>>>> Therefore, at the end of rcu_gp_init(), immediately report a quiescent
> > >>>>> state for the GP kthread's CPU using rcu_qs() + rcu_report_qs_rdp(). The
> > >>>>> GP kthread cannot be in an RCU read-side critical section while running
> > >>>>> GP initialization, so this is safe and results in significant latency
> > >>>>> improvements.
> > >>>>> 
> > >>>>> I benchmarked 100 synchronize_rcu() calls with 32 CPUs, 10 runs each
> > >>>>> showing significant latency improvements (default settings for fqs jiffies):
> > >>>>> 
> > >>>>> Baseline (without fix):
> > >>>>> | Run | Mean      | Min      | Max       |
> > >>>>> |-----|-----------|----------|-----------|
> > >>>>> | 1   | 10.088 ms | 9.989 ms | 18.848 ms |
> > >>>>> | 2   | 10.064 ms | 9.982 ms | 16.470 ms |
> > >>>>> | 3   | 10.051 ms | 9.988 ms | 15.113 ms |
> > >>>>> | 4   | 10.125 ms | 9.929 ms | 22.411 ms |
> > >>>>> | 5   |  8.695 ms | 5.996 ms | 15.471 ms |
> > >>>>> | 6   | 10.157 ms | 9.977 ms | 25.723 ms |
> > >>>>> | 7   | 10.102 ms | 9.990 ms | 20.224 ms |
> > >>>>> | 8   |  8.050 ms | 5.985 ms | 10.007 ms |
> > >>>>> | 9   | 10.059 ms | 9.978 ms | 15.934 ms |
> > >>>>> | 10  | 10.077 ms | 9.984 ms | 17.703 ms |
> > >>>>> 
> > >>>>> With fix:
> > >>>>> | Run | Mean     | Min      | Max       |
> > >>>>> |-----|----------|----------|-----------|
> > >>>>> | 1   | 6.027 ms | 5.915 ms |  8.589 ms |
> > >>>>> | 2   | 6.032 ms | 5.984 ms |  9.241 ms |
> > >>>>> | 3   | 6.010 ms | 5.986 ms |  7.004 ms |
> > >>>>> | 4   | 6.076 ms | 5.993 ms | 10.001 ms |
> > >>>>> | 5   | 6.084 ms | 5.893 ms | 10.250 ms |
> > >>>>> | 6   | 6.034 ms | 5.908 ms |  9.456 ms |
> > >>>>> | 7   | 6.051 ms | 5.993 ms | 10.000 ms |
> > >>>>> | 8   | 6.057 ms | 5.941 ms | 10.001 ms |
> > >>>>> | 9   | 6.016 ms | 5.927 ms |  7.540 ms |
> > >>>>> | 10  | 6.036 ms | 5.993 ms |  9.579 ms |
> > >>>>> 
> > >>>>> Summary:
> > >>>>> - Mean latency: 9.75 ms -> 6.04 ms (38% improvement)
> > >>>>> - Max latency:  25.72 ms -> 10.25 ms (60% improvement)
> > >>>>> 
> > >>>>> Tested rcutorture TREE and SRCU configurations.
> > >>>>> 
> > >>>>> [apply paulmck feedack on moving logic to rcu_gp_init()]
> > >>>> 
> > >>>> If anything, these numbers look better, so good show!!!
> > >>> 
> > >>> Thanks, I ended up collecting more samples in the v2 to further confirm the
> > >>> improvements.
> > >>> 
> > >>>> Are there workloads that might be hurt by some side effect such
> > >>>> as increased CPU utilization by the RCU grace-period kthread?  One
> > >>>> non-mainstream hypothetical situation that comes to mind is a kernel
> > >>>> built with SMP=y but running on a single-CPU system with a high-frequence
> > >>>> periodic interrupt that does call_rcu().  Might that result in the RCU
> > >>>> grace-period kthread chewing up the entire CPU?
> > >>> 
> > >>> There are still GP delays due to FQS, even with this change, so it could not
> > >>> chew up the entire CPU I believe. The GP cycle should still insert delays
> > >>> into the GP kthread. I did not notice in my testing that synchronize_rcu()
> > >>> latency dropping to sub millisecond, it was still limited by the timer wheel
> > >>> delays and the FQS delays.
> > >>> 
> > >>>> For a non-hypothetical case, could you please see if one of the
> > >>>> battery-powered embedded guys would be willing to test this?
> > >>> 
> > >>> My suspicion is the battery-powered folks are already running RCU_LAZY to
> > >>> reduce RCU activity, so they wouldn't be effected. call_rcu() during idleness
> > >>> will be going to the bypass. Last I checked, Android and ChromeOS were both
> > >>> enabling RCU_LAZY everywhere (back when I was at Google).
> > >>> 
> > >>> Uladzislau works on embedded (or at least till recently) and had recently
> > >>> checked this area for improvements so I think he can help quantify too
> > >>> perhaps. He is on CC. I personally don't directly work on embedded at the
> > >>> moment, just big compute hungry machines. ;-) Uladzislau, would you have some
> > >>> time to test on your Android devices?
> > >>> 
> > >> I will check the patch on my home based systems, big machines also :)
> > >> I do not work with mobile area any more thus do not have access to our
> > >> mobile devices. In fact i am glad that i have switched to something new.
> > >> I was a bit tired by the applied Google restrictions when it comes to
> > >> changes to the kernel and other Android layers.
> > > 
> > > How quickly I forget!  ;-)
> > > 
> > > Any thoughts on who would be a good person to ask about testing Joel's
> > > patch on mobile platforms?
> > 
> > Maybe Suren? As precedent and fwiw, When rcu_normal_wake_from_gp
> > optimization happened, it only improved things for Android.
> > 
> > Also Android already uses RCU_LAZY so this should not affect power for
> > non-hurry usages.
> > 
> > Also networking bridge removal depends on synchronize_rcu() latency. When
> > I forced rcu_normal_wake_from_gp on large machines, it improved bridge
> > removal speed by about 5% per my notes. I would expect similar
> > improvements with this.
> 
> Could you please try running on a single-CPU system or VM to check the
> CPU overhead from RCU's grace-period kthread?

Hi, Paul,

I ran some tests with single CPU and used perf to measure overhead of the
GP kthread (rcu_preempt).

Actually, the GP kthread's CPU usage goes down, I believe this is because it
sleeps more.

I see similar/same results with synchronize_rcu() loop (200 iterations). I
also tested call_rcu() stressing from timer interrupts and call_rcu_hurry()
flooding.

                          Baseline    With Patch    Change
  task-clock:             1008 ms     898 ms        -11%
  CPU cycles:             48M         44M           -8%

Should I add these results to the changelog and send out a v3 (preferrably
with your review tag if you approve).

thanks,

 - Joel