Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Subject: Re: [PATCH] prctl_set_mm: downgrade mmap_sem to read lock
To:     =?UTF-8?Q?Michal_Koutn=c3=bd?= <mkoutny@suse.com>,
        Cyrill Gorcunov <gorcunov@gmail.com>
Cc:     mhocko@kernel.org, akpm@linux-foundation.org,
        arunks@codeaurora.org, brgl@bgdev.pl, geert+renesas@glider.be,
        linux-kernel@vger.kernel.org, linux-mm@kvack.org,
        mguzik@redhat.com, rppt@linux.ibm.com, vbabka@suse.cz
References: <20190417145548.GN5878@dhcp22.suse.cz>
 <20190418135039.19987-1-mkoutny@suse.com>
From:   Laurent Dufour <ldufour@linux.ibm.com>
Date:   Thu, 18 Apr 2019 16:27:53 +0200
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.14; rv:60.0)
 Gecko/20100101 Thunderbird/60.6.1
MIME-Version: 1.0
In-Reply-To: <20190418135039.19987-1-mkoutny@suse.com>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: en-US
Content-Transfer-Encoding: 8bit
Message-Id: <27defd37-7e4e-f919-fe0c-64e1efdafdcf@linux.ibm.com>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Le 18/04/2019 à 15:50, Michal Koutný a écrit :
> I learnt, it's, alas, too late to drop the non PRCTL_SET_MM_MAP calls
> [1], so at least downgrade the write acquisition of mmap_sem as in the
> patch below (that should be stacked on the previous one or squashed).
> 
> Cyrill, you mentioned lock changes in [1] but the link seems empty. Is
> it supposed to be [2]? That could be an alternative to this patch after
> some refreshments and clarifications.
> 
> 
> [1] https://lore.kernel.org/lkml/20190417165632.GC3040@uranus.lan/
> [2] https://lore.kernel.org/lkml/20180507075606.870903028@gmail.com/
> 
> ========
> 
> Since commit 88aa7cc688d4 ("mm: introduce arg_lock to protect
> arg_start|end and env_start|end in mm_struct") we use arg_lock for
> boundaries modifications. Synchronize prctl_set_mm with this lock and
> keep mmap_sem for reading only (analogous to what we already do in
> prctl_set_mm_map).
> 
> Also, save few cycles by looking up VMA only after performing basic
> arguments validation.
> 
> Signed-off-by: Michal Koutný <mkoutny@suse.com>
> ---
>   kernel/sys.c | 12 +++++++++---
>   1 file changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sys.c b/kernel/sys.c
> index 12df0e5434b8..bbce0f26d707 100644
> --- a/kernel/sys.c
> +++ b/kernel/sys.c
> @@ -2125,8 +2125,12 @@ static int prctl_set_mm(int opt, unsigned long addr,
>   
>   	error = -EINVAL;
>   
> -	down_write(&mm->mmap_sem);
> -	vma = find_vma(mm, addr);
> +	/*
> +	 * arg_lock protects concurent updates of arg boundaries, we need mmap_sem for
> +	 * a) concurrent sys_brk, b) finding VMA for addr validation.
> +	 */
> +	down_read(&mm->mmap_sem);
> +	spin_lock(&mm->arg_lock);
>   
>   	prctl_map.start_code	= mm->start_code;
>   	prctl_map.end_code	= mm->end_code;
> @@ -2185,6 +2189,7 @@ static int prctl_set_mm(int opt, unsigned long addr,
>   	if (error)
>   		goto out;
>   
> +	vma = find_vma(mm, addr);

Why is find_vma() called while holding the arg_lock ?

To limit the time the spinlock is held, would it be better to
    	read_lock(mmap_sem)
    	find_vma()
    	spin_lock(arg_lock)
    	..
out:
	spin_unlock()
	up_read(mmap_sem)

Not sure this would change a lot the performance anyway.

>   	switch (opt) {
>   	/*
>   	 * If command line arguments and environment
> @@ -2218,7 +2223,8 @@ static int prctl_set_mm(int opt, unsigned long addr,
>   
>   	error = 0;
>   out:
> -	up_write(&mm->mmap_sem);
> +	spin_unlock(&mm->arg_lock);
> +	up_read(&mm->mmap_sem);
>   	return error;
>   }
>   
>