Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20221109203051.1835763-1-torvalds@linux-foundation.org>
 <20221109203051.1835763-4-torvalds@linux-foundation.org> <CAHk-=wjf+gN25grUT3o3XK8-B-b2jhBuN8YMLQvq-=AXTcuFXg@mail.gmail.com>
 <CAHk-=wiFmm+X92Ghkz_LDG53djReK=idAj0uvSdES+yeG1X=Dw@mail.gmail.com>
 <Y3SVqePHGXRUagyF@li-4a3a4a4c-28e5-11b2-a85c-a8d192c6f089.ibm.com> <Y3SWTXMotahiqvBO@li-4a3a4a4c-28e5-11b2-a85c-a8d192c6f089.ibm.com>
In-Reply-To: <Y3SWTXMotahiqvBO@li-4a3a4a4c-28e5-11b2-a85c-a8d192c6f089.ibm.com>
From:   Linus Torvalds <torvalds@linux-foundation.org>
Date:   Wed, 16 Nov 2022 09:52:06 -0800
Message-ID: <CAHk-=wh-oBpkN8AKALDkeTcVRzkUqqUyFevCNcCy3F76ogGonA@mail.gmail.com>
Subject: Re: mm: mmu_gather: do not define delayed_rmap if not used
To:     Alexander Gordeev <agordeev@linux.ibm.com>
Cc:     Hugh Dickins <hughd@google.com>,
        Johannes Weiner <hannes@cmpxchg.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org,
        Nadav Amit <nadav.amit@gmail.com>,
        Will Deacon <will@kernel.org>,
        Aneesh Kumar <aneesh.kumar@linux.ibm.com>,
        Nick Piggin <npiggin@gmail.com>,
        Heiko Carstens <hca@linux.ibm.com>,
        Vasily Gorbik <gor@linux.ibm.com>,
        Christian Borntraeger <borntraeger@linux.ibm.com>,
        Sven Schnelle <svens@linux.ibm.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Nov 15, 2022 at 11:51 PM Alexander Gordeev
<agordeev@linux.ibm.com> wrote:
>
> In cases the delayed rmap removal is not used (which are
> currently UP and s390) skip delayed_rmap flag and make
> the related code paths no-op.

So I'm not convinced about this patch.

I particularly dislike adding even more #ifdef's around the data
structure - it already is pretty nasty, and it was hard to see where
things were initialized.

The only actual code impact of this is in tlb_next_batch(), which
tests for "do I have delayed rmaps pending, in which case I won't add
new batches". Everything else is already either optimized away, or
just "one bit declared in a structure that already has bitfields and
has room for several extra bits":

And that "I need to allocate new batches" case really doesn't matter
anyway - it's not even build at all on s390, and on UP where it's
there but technically pointless to have the test it really isn't
noticeable.

So the previous patch I was "this shouldn't actually _matter_, but it
does seem cleaner to do it this way".

But _this_ patch makes me go "it still doesn't matter, but now this
patch is actually adding extra infrastructure for the 'not-mattering'
case".

So I don't _hate_ this patch, but I think this actually makes the
current mess wrt our 'struct mmu_gather' worse rather than better.

That structure is already a pain, with horrendous initialization and
different bit-fields having different lifetimes. I'd rather have one
unconditional simple bitfield, than have another bitfield that has
conditional complications.

              Linus