Received-SPF: pass (google.com: domain of linux-kernel+bounces-199806-linux.lists.archive=gmail.com@vger.kernel.org designates 139.178.88.99 as permitted sender) client-ip=139.178.88.99;
Date: Mon, 3 Jun 2024 16:03:05 -0700
In-Reply-To: <CADrL8HW44Hx_Ejx_6+FVKt1V17PdgT6rw+sNtKzumqc9UCVDfA@mail.gmail.com>
Precedence: bulk
Mime-Version: 1.0
References: <20240529180510.2295118-1-jthoughton@google.com>
 <20240529180510.2295118-3-jthoughton@google.com> <CAOUHufYFHKLwt1PWp2uS6g174GZYRZURWJAmdUWs5eaKmhEeyQ@mail.gmail.com>
 <ZlelW93_T6P-ZuSZ@google.com> <CAOUHufZdEpY6ra73SMHA33DegKxKaUM=Os7A7aDBFND6NkbUmQ@mail.gmail.com>
 <Zley-u_dOlZ-S-a6@google.com> <CADrL8HXHWg_MkApYQTngzmN21NEGNWC6KzJDw_Lm63JHJkR=5A@mail.gmail.com>
 <CAOUHufZq6DwpStzHtjG+TOiHaQ6FFbkTfHMCe8Yy0n_M9MKdqw@mail.gmail.com> <CADrL8HW44Hx_Ejx_6+FVKt1V17PdgT6rw+sNtKzumqc9UCVDfA@mail.gmail.com>
Message-ID: <Zl5LqcusZ88QOGQY@google.com>
Subject: Re: [PATCH v4 2/7] mm: multi-gen LRU: Have secondary MMUs participate
 in aging
From: Sean Christopherson <seanjc@google.com>
To: James Houghton <jthoughton@google.com>
Cc: Yu Zhao <yuzhao@google.com>, Andrew Morton <akpm@linux-foundation.org>, 
	Paolo Bonzini <pbonzini@redhat.com>, Albert Ou <aou@eecs.berkeley.edu>, 
	Ankit Agrawal <ankita@nvidia.com>, Anup Patel <anup@brainfault.org>, 
	Atish Patra <atishp@atishpatra.org>, Axel Rasmussen <axelrasmussen@google.com>, 
	Bibo Mao <maobibo@loongson.cn>, Catalin Marinas <catalin.marinas@arm.com>, 
	David Matlack <dmatlack@google.com>, David Rientjes <rientjes@google.com>, 
	Huacai Chen <chenhuacai@kernel.org>, James Morse <james.morse@arm.com>, 
	Jonathan Corbet <corbet@lwn.net>, Marc Zyngier <maz@kernel.org>, Michael Ellerman <mpe@ellerman.id.au>, 
	Nicholas Piggin <npiggin@gmail.com>, Oliver Upton <oliver.upton@linux.dev>, 
	Palmer Dabbelt <palmer@dabbelt.com>, Paul Walmsley <paul.walmsley@sifive.com>, 
	Raghavendra Rao Ananta <rananta@google.com>, Ryan Roberts <ryan.roberts@arm.com>, 
	Shaoqin Huang <shahuang@redhat.com>, Shuah Khan <shuah@kernel.org>, 
	Suzuki K Poulose <suzuki.poulose@arm.com>, Tianrui Zhao <zhaotianrui@loongson.cn>, 
	Will Deacon <will@kernel.org>, Zenghui Yu <yuzenghui@huawei.com>, kvm-riscv@lists.infradead.org, 
	kvm@vger.kernel.org, kvmarm@lists.linux.dev, 
	linux-arm-kernel@lists.infradead.org, linux-doc@vger.kernel.org, 
	linux-kernel@vger.kernel.org, linux-kselftest@vger.kernel.org, 
	linux-mips@vger.kernel.org, linux-mm@kvack.org, 
	linux-riscv@lists.infradead.org, linuxppc-dev@lists.ozlabs.org, 
	loongarch@lists.linux.dev
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: quoted-printable

On Mon, Jun 03, 2024, James Houghton wrote:
> On Thu, May 30, 2024 at 11:06=E2=80=AFPM Yu Zhao <yuzhao@google.com> wrot=
e:
> > What I don't think is acceptable is simplifying those optimizations
> > out without documenting your justifications (I would even call it a
> > design change, rather than simplification, from v3 to v4).
>=20
> I'll put back something similar to what you had before (like a
> test_clear_young() with a "fast" parameter instead of "bitmap"). I
> like the idea of having a new mmu notifier, like
> fast_test_clear_young(), while leaving test_young() and clear_young()
> unchanged (where "fast" means "prioritize speed over accuracy").

Those two statements are contradicting each other, aren't they?  Anyways, I=
 vote
for a "fast only" variant, e.g. test_clear_young_fast_only() or so.  gup() =
has
already established that terminology in mm/, so hopefully it would be famil=
iar
to readers.  We could pass a param, but then the MGLRU code would likely en=
d up
doing a bunch of useless indirect calls into secondary MMUs, whereas a dedi=
cated
hook allows implementations to nullify the pointer if the API isn't support=
ed
for whatever reason.

And pulling in Oliver's comments about locking, I think it's important that=
 the
mmu_notifier API express it's requirement that the operation be "fast", not=
 that
it be lockless.  E.g. if a secondary MMU can guarantee that a lock will be
contented only in rare, slow cases, then taking a lock is a-ok.  Or a secon=
dary
MMU could do try-lock and bail if the lock is contended.

That way KVM can honor the intent of the API with an implementation that wo=
rks
best for KVM _and_ for MGRLU.  I'm sure there will be future adjustments an=
d fixes,
but that's just more motivation for using something like "fast only" instea=
d of
"lockless".

> > > I made this logic change as part of removing batching.
> > >
> > > I'd really appreciate guidance on what the correct thing to do is.
> > >
> > > In my mind, what would work great is: by default, do aging exactly
> > > when KVM can do it locklessly, and then have a Kconfig to always have
> > > MGLRU to do aging with KVM if a user really cares about proactive
> > > reclaim (when the feature bit is set). The selftest can check the
> > > Kconfig + feature bit to know for sure if aging will be done.
> >
> > I still don't see how that Kconfig helps. Or why the new static branch
> > isn't enough?
>=20
> Without a special Kconfig, the feature bit just tells us that aging
> with KVM is possible, not that it will necessarily be done. For the
> self-test, it'd be good to know exactly when aging is being done or
> not, so having a Kconfig like LRU_GEN_ALWAYS_WALK_SECONDARY_MMU would
> help make the self-test set the right expectations for aging.
>=20
> The Kconfig would also allow a user to know that, no matter what,
> we're going to get correct age data for VMs, even if, say, we're using
> the shadow MMU.

Heh, unless KVM flushes, you won't get "correct" age data.

> This is somewhat important for me/Google Cloud. Is that reasonable? Maybe
> there's a better solution.

Hmm, no?  There's no reason to use a Kconfig, e.g. if we _really_ want to p=
rioritize
accuracy over speed, then a KVM (x86?) module param to have KVM walk nested=
 TDP
page tables would give us what we want.

But before we do that, I think we need to perform due dilegence (or provide=
 data)
showing that having KVM take mmu_lock for write in the "fast only" API prov=
ides
better total behavior.  I.e. that the additional accuracy is indeed worth t=
he cost.