Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210630051118.2212-1-yajun.deng@linux.dev>
In-Reply-To: <20210630051118.2212-1-yajun.deng@linux.dev>
From:   Eric Dumazet <edumazet@google.com>
Date:   Wed, 30 Jun 2021 13:02:53 +0200
Message-ID: <CANn89iJp8NoGeqP47u22tXpO78gzQoJBQQ4dpSdGb+v_dZ9kPQ@mail.gmail.com>
Subject: Re: [PATCH] net: core: Modify alloc_size in alloc_netdev_mqs()
To:     Yajun Deng <yajun.deng@linux.dev>
Cc:     davem@davemloft.net, kuba@kernel.org, andriin@fb.com,
        atenart@kernel.org, alobakin@pm.me, ast@kernel.org,
        daniel@iogearbox.net, weiwan@google.com, ap420073@gmail.com,
        netdev@vger.kernel.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Wed, Jun 30, 2021 at 7:11 AM Yajun Deng <yajun.deng@linux.dev> wrote:
>
> Use ALIGN for 'struct net_device', and remove the unneeded
> 'NETDEV_ALIGN - 1'. This can save a few bytes. and modify
> the pr_err content when txqs < 1.

I think that in old times (maybe still today), SLAB debugging could
lead to not unaligned allocated zones.
The forced alignment for netdev structures came in
commit f346af6a27c0cea99522213cb813fd30489136e2 ("net_device and
netdev private struct allocation improvements.")
in linux-2.6.3 (back in 2004)

This supposedly was a win in itself, otherwise Al Viro would not have
spent time on this.

>
> Signed-off-by: Yajun Deng <yajun.deng@linux.dev>
> ---
>  net/core/dev.c | 10 ++++------
>  1 file changed, 4 insertions(+), 6 deletions(-)
>
> diff --git a/net/core/dev.c b/net/core/dev.c
> index c253c2aafe97..c42a682a624d 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -10789,7 +10789,7 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
>         BUG_ON(strlen(name) >= sizeof(dev->name));
>
>         if (txqs < 1) {
> -               pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
> +               pr_err("alloc_netdev: Unable to allocate device with zero TX queues\n");
>                 return NULL;
>         }
>
> @@ -10798,14 +10798,12 @@ struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
>                 return NULL;
>         }
>
> -       alloc_size = sizeof(struct net_device);
> +       /* ensure 32-byte alignment of struct net_device*/
> +       alloc_size = ALIGN(sizeof(struct net_device), NETDEV_ALIGN);

This is not really needed, because struct net_device is cache line
aligned already on SMP builds.

>         if (sizeof_priv) {
>                 /* ensure 32-byte alignment of private area */
> -               alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
> -               alloc_size += sizeof_priv;
> +               alloc_size += ALIGN(sizeof_priv, NETDEV_ALIGN);

No longer needed, the private area starts at the end of struct net_device, whose
size is a multiple of cache line.

Really I doubt this makes sense anymore these days, we have hundreds
of structures in the kernel
that would need a similar handling if SLAB/SLUB was doing silly things.

I would simply do :

alloc_size += sizeof_priv;


>         }
> -       /* ensure 32-byte alignment of whole construct */
> -       alloc_size += NETDEV_ALIGN - 1;
>
>         p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
>         if (!p)
> --
> 2.32.0
>