[PATCH 39/41] kernel/fork: throttle call_rcu() calls in vm_area_free

[Paul E. McKenney]

On Fri, Jan 20, 2023 at 09:57:05AM +0100, Michal Hocko wrote:
> On Thu 19-01-23 11:17:07, Paul E. McKenney wrote:
> > On Thu, Jan 19, 2023 at 01:52:14PM +0100, Michal Hocko wrote:
> > > On Wed 18-01-23 11:01:08, Suren Baghdasaryan wrote:
> > > > On Wed, Jan 18, 2023 at 10:34 AM Paul E. McKenney <paulmck at kernel.org> wrote:
> > > [...]
> > > > > There are a couple of possibilities here.
> > > > >
> > > > > First, if I am remembering correctly, the time between the call_rcu()
> > > > > and invocation of the corresponding callback was taking multiple seconds,
> > > > > but that was because the kernel was built with CONFIG_LAZY_RCU=y in
> > > > > order to save power by batching RCU work over multiple call_rcu()
> > > > > invocations.  If this is causing a problem for a given call site, the
> > > > > shiny new call_rcu_hurry() can be used instead.  Doing this gets back
> > > > > to the old-school non-laziness, but can of course consume more power.
> > > > 
> > > > That would not be the case because CONFIG_LAZY_RCU was not an option
> > > > at the time I was profiling this issue.
> > > > Laxy RCU would be a great option to replace this patch but
> > > > unfortunately it's not the default behavior, so I would still have to
> > > > implement this batching in case lazy RCU is not enabled.
> > > > 
> > > > >
> > > > > Second, there is a much shorter one-jiffy delay between the call_rcu()
> > > > > and the invocation of the corresponding callback in kernels built with
> > > > > either CONFIG_NO_HZ_FULL=y (but only on CPUs mentioned in the nohz_full
> > > > > or rcu_nocbs kernel boot parameters) or CONFIG_RCU_NOCB_CPU=y (but only
> > > > > on CPUs mentioned in the rcu_nocbs kernel boot parameters).  The purpose
> > > > > of this delay is to avoid lock contention, and so this delay is incurred
> > > > > only on CPUs that are queuing callbacks at a rate exceeding 16K/second.
> > > > > This is reduced to a per-jiffy limit, so on a HZ=1000 system, a CPU
> > > > > invoking call_rcu() at least 16 times within a given jiffy will incur
> > > > > the added delay.  The reason for this delay is the use of a separate
> > > > > ->nocb_bypass list.  As Suren says, this bypass list is used to reduce
> > > > > lock contention on the main ->cblist.  This is not needed in old-school
> > > > > kernels built without either CONFIG_NO_HZ_FULL=y or CONFIG_RCU_NOCB_CPU=y
> > > > > (including most datacenter kernels) because in that case the callbacks
> > > > > enqueued by call_rcu() are touched only by the corresponding CPU, so
> > > > > that there is no need for locks.
> > > > 
> > > > I believe this is the reason in my profiled case.
> > > > 
> > > > >
> > > > > Third, if you are instead seeing multiple milliseconds of CPU consumed by
> > > > > call_rcu() in the common case (for example, without the aid of interrupts,
> > > > > NMIs, or SMIs), please do let me know.  That sounds to me like a bug.
> > > > 
> > > > I don't think I've seen such a case.
> > > > Thanks for clarifications, Paul!
> > > 
> > > Thanks for the explanation Paul. I have to say this has caught me as a
> > > surprise. There are just not enough details about the benchmark to
> > > understand what is going on but I find it rather surprising that
> > > call_rcu can induce a higher overhead than the actual kmem_cache_free
> > > which is the callback. My naive understanding has been that call_rcu is
> > > really fast way to defer the execution to the RCU safe context to do the
> > > final cleanup.
> > 
> > If I am following along correctly (ha!), then your "induce a higher
> > overhead" should be something like "induce a higher to-kfree() latency".
> 
> Yes, this is expected.
> 
> > Of course, there already is a higher latency-to-kfree via call_rcu()
> > than via a direct call to kfree(), and callback-offload CPUs that are
> > being flooded with callbacks raise that latency a jiffy or so more in
> > order to avoid lock contention.
> > 
> > If this becomes a problem, the callback-offloading code can be a bit
> > smarter about avoiding lock contention, but need to see a real problem
> > before I make that change.  But if there is a real problem I will of
> > course fix it.
> 
> I believe that Suren claims that the call_rcu is really visible in the
> exit_mmap case. Time-to-free actual vmas shouldn't really be material
> for that path. If that happens much more later on there could be some
> side effects by an increased memory consumption but that should be
> marginal. How fast exit_mmap really is should only depend on direct
> calls from that path.
> 
> But I guess we need some specific numbers from Suren to be sure what is
> going on here.

Actually, Suren did discuss these (perhaps offlist) back in August.
I was just being forgetful.  :-/

							Thanx, Paul