Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:8 as permitted sender) client-ip=2620:137:e000::3:8;
MIME-Version: 1.0
References: <20231207094844.706407-1-agordeev@linux.ibm.com>
In-Reply-To: <20231207094844.706407-1-agordeev@linux.ibm.com>
From:   Jann Horn <jannh@google.com>
Date:   Thu, 7 Dec 2023 15:00:10 +0100
Message-ID: <CAG48ez2KS=EpmRmiQWB0jOMD+qCGjZag6dtFQQG5FUfWU0RfUw@mail.gmail.com>
Subject: Re: [PATCH] mm: do not protect VMA lock object in vma_end_read()
To:     Alexander Gordeev <agordeev@linux.ibm.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org,
        Suren Baghdasaryan <surenb@google.com>,
        Waiman Long <longman@redhat.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>,
        Thomas Gleixner <tglx@linutronix.de>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, Dec 7, 2023 at 10:49=E2=80=AFAM Alexander Gordeev
<agordeev@linux.ibm.com> wrote:
> In early discussion to the implementation of vma_end_read()
> Jann Horn pointed out that up_read() could access the VMA
> lock object after it has already been acquired by someone
> else. As result, up_read() is protected with RCU read lock:
>
>         rcu_read_lock(); /* keeps vma alive */
>         up_read(&vma->lock);
>         rcu_read_unlock();
>
> Since commit 3f5245538a19 ("locking/rwsem: Disable preemption
> in all down_read*() and up_read() code paths") __up_read()
> disables preemption internally and thus the need to protect
> the VMA lock object does not exist anymore.

I think this is a bad idea. Please don't.

Yes, it looks like the (non-RT) implementation of __up_read
*currently* disables preemption. But that's an implementation detail,
not a documented API contract of up_read(), so there would be nothing
stopping someone from reimplementing __up_read() in the future such
that the preemption stuff disappears again.

And from what I can tell from a quick look, the RT implementation of
__up_read() does not currently give you this kind of guarantee.

In my opinion, if you want to make this change, then as a prerequisite
you have to get buy-in from the locking maintainers.


> Link: https://lore.kernel.org/all/CAG48ez3sCwasFzKD5CsqMFA2W57-2fazd75g7r=
0NaA_BVNTLow@mail.gmail.com/
> Cc: Jann Horn <jannh@google.com>
> Cc: Suren Baghdasaryan <surenb@google.com>
> Signed-off-by: Alexander Gordeev <agordeev@linux.ibm.com>
> ---
>  include/linux/mm.h | 2 --
>  1 file changed, 2 deletions(-)
>
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 418d26608ece..7b32bc75a4ab 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -683,9 +683,7 @@ static inline bool vma_start_read(struct vm_area_stru=
ct *vma)
>
>  static inline void vma_end_read(struct vm_area_struct *vma)
>  {
> -       rcu_read_lock(); /* keeps vma alive till the end of up_read */
>         up_read(&vma->vm_lock->lock);
> -       rcu_read_unlock();
>  }
>
>  /* WARNING! Can only be used if mmap_lock is expected to be write-locked=
 */
> --
> 2.40.1
>