From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jason Wang Subject: Re: [net-next rfc 1/3] net: avoid high order memory allocation for queues by using flex array Date: Thu, 20 Jun 2013 13:14:15 +0800 Message-ID: <51C28FA7.2070901@redhat.com> References: <1371620452-49349-1-git-send-email-jasowang@redhat.com> <1371620452-49349-2-git-send-email-jasowang@redhat.com> <1371623518.3252.267.camel@edumazet-glaptop> <20130619091132.GA2816@redhat.com> <1371635763.3252.289.camel@edumazet-glaptop> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Cc: "Michael S. Tsirkin" , davem@davemloft.net, edumazet@google.com, hkchu@google.com, netdev@vger.kernel.org, linux-kernel@vger.kernel.org To: Eric Dumazet Return-path: In-Reply-To: <1371635763.3252.289.camel@edumazet-glaptop> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On 06/19/2013 05:56 PM, Eric Dumazet wrote: > On Wed, 2013-06-19 at 12:11 +0300, Michael S. Tsirkin wrote: > >> Well KVM supports up to 160 VCPUs on x86. >> >> Creating a queue per CPU is very reasonable, and >> assuming cache line size of 64 bytes, netdev_queue seems to be 320 >> bytes, that's 320*160 = 51200. So 12.5 pages, order-4 allocation. >> I agree most people don't have such systems yet, but >> they do exist. > Even so, it will just work, like a fork() is likely to work, even if a > process needs order-1 allocation for kernel stack. > > Some drivers still use order-10 allocations with kmalloc(), and nobody > complained yet. > > We had complains with mlx4 driver lately only bcause kmalloc() now gives > a warning if allocations above MAX_ORDER are attempted. > > Having a single pointer means that we can : > > - Attempts a regular kmalloc() call, it will work most of the time. > - fallback to vmalloc() _if_ kmalloc() failed. > > Frankly, if you want one tx queue per cpu, I would rather use > NETIF_F_LLTX, like some other virtual devices. A drawback of NETIF_F_LLTX is that we may contend on qdisc lock especially when we have a huge number of tx queues. > > This way, you can have real per cpu memory, with proper NUMA affinity. > > >