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
In-Reply-To: <CAHk-=wiG9xaVvBJXHqTxtop0=mW9KxPS9C54ED23p59VNEKdWg@mail.gmail.com>
References: <20230708191212.4147700-1-surenb@google.com> <20230708191212.4147700-3-surenb@google.com>
 <20230804214620.btgwhsszsd7rh6nf@f> <CAHk-=wiy125k1dBmQFTGpHwiOqEyrD6xnd4xKWfe97H_HodgDA@mail.gmail.com>
 <CAGudoHFsAU_BDCOuz8UgDBLGEM8xg=aUGjaVoqkM_Zvxo2Re_g@mail.gmail.com> <CAHk-=wiG9xaVvBJXHqTxtop0=mW9KxPS9C54ED23p59VNEKdWg@mail.gmail.com>
From:   Mateusz Guzik <mjguzik@gmail.com>
Date:   Sat, 5 Aug 2023 03:06:13 +0200
Message-ID: <CAGudoHFsxmiezZREyVkrPYBBij3u9SNaxLWipOOPy5S+5pJcZA@mail.gmail.com>
Subject: Re: [PATCH v2 3/3] fork: lock VMAs of the parent process when forking
To:     Linus Torvalds <torvalds@linux-foundation.org>
Cc:     Suren Baghdasaryan <surenb@google.com>, akpm@linux-foundation.org,
        regressions@leemhuis.info, bagasdotme@gmail.com,
        jacobly.alt@gmail.com, willy@infradead.org,
        liam.howlett@oracle.com, david@redhat.com, peterx@redhat.com,
        ldufour@linux.ibm.com, linux-kernel@vger.kernel.org,
        linux-mm@kvack.org, linuxppc-dev@lists.ozlabs.org,
        linux-arm-kernel@lists.infradead.org, gregkh@linuxfoundation.org,
        regressions@lists.linux.dev, Jiri Slaby <jirislaby@kernel.org>,
        =?UTF-8?Q?Holger_Hoffst=C3=A4tte?= <holger@applied-asynchrony.com>,
        stable@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On 8/5/23, Linus Torvalds <torvalds@linux-foundation.org> wrote:
> On Fri, 4 Aug 2023 at 16:25, Mateusz Guzik <mjguzik@gmail.com> wrote:
>>
>> I know of these guys, I think they are excluded as is -- they go
>> through access_remote_vm, starting with:
>>         if (mmap_read_lock_killable(mm))
>>                 return 0;
>>
>> while dup_mmap already write locks the parent's mm.
>
> Oh, you're only worried about vma_start_write()?
>
> That's a non-issue. It doesn't take the lock normally, since it starts off
> with
>
>         if (__is_vma_write_locked(vma, &mm_lock_seq))
>                 return;
>
> which catches on the lock sequence number already being set.
>
> So no extra locking there.
>
> Well, technically there's extra locking because the code stupidly
> doesn't initialize new vma allocations to the right sequence number,
> but that was talked about here:
>
>
> https://lore.kernel.org/all/CAHk-=wiCrWAoEesBuoGoqqufvesicbGp3cX0LyKgEvsFaZNpDA@mail.gmail.com/
>
> and it's a separate issue.
>

I'm going to bet one beer this is the issue.

The patch I'm responding to only consists of adding the call to
vma_start_write and claims the 5% slowdown from it, while fixing
crashes if the forking process is multithreaded.

For the fix to work it has to lock something against the parent.

        VMA_ITERATOR(old_vmi, oldmm, 0);
[..]
        for_each_vma(old_vmi, mpnt) {
[..]
                vma_start_write(mpnt);

the added line locks an obj in the parent's vm space.

The problem you linked looks like pessimization for freshly allocated
vmas, but that's what is being operated on here.

-- 
Mateusz Guzik <mjguzik gmail.com>