From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jason Wang Subject: Re: [PATCH] net-tun: restructure tun_do_read for better sleep/wakeup efficiency Date: Fri, 09 May 2014 11:10:43 +0800 Message-ID: <536C4733.9020704@redhat.com> References: <1399422244-22751-1-git-send-email-xii@google.com> <5369AB36.6030609@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: "David S. Miller" , netdev@vger.kernel.org, Maxim Krasnyansky , Neal Cardwell , Eric Dumazet To: Xi Wang Return-path: Received: from mx1.redhat.com ([209.132.183.28]:21259 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755500AbaEIDKp (ORCPT ); Thu, 8 May 2014 23:10:45 -0400 In-Reply-To: Sender: netdev-owner@vger.kernel.org List-ID: On 05/09/2014 02:22 AM, Xi Wang wrote: > On Tue, May 6, 2014 at 8:40 PM, Jason Wang wrote: >> On 05/07/2014 08:24 AM, Xi Wang wrote: >>> tun_do_read always adds current thread to wait queue, even if a packet >>> is ready to read. This is inefficient because both sleeper and waker >>> want to acquire the wait queue spin lock when packet rate is high. >> After commit 61a5ff15ebdab87887861a6b128b108404e4706d, this will only >> help for blocking read. Looks like for performance critical userspaces, >> they will use non blocking reads. >>> We restructure the read function and use common kernel networking >>> routines to handle receive, sleep and wakeup. With the change >>> available packets are checked first before the reading thread is added >>> to the wait queue. >> This is interesting, since it may help if we want to add rx busy loop >> for tun. (In fact I worked a similar patch like this). > > Yes this should be a good side effect and I am also interested in trying. > Busy polling in user space is not ideal as it doesn't give the lowest latency. > Besides differences in interrupt latency etc., there is a bad case for > non blocking mode: When a packet arrives right before the polling thread > returns to userspace. The control flow has to cross kernel/userspace > boundary 3 times before the packet can be processed, while kernel > blocking or busy polling only needs 1 boundary crossing. So if we want to implement this, we need a feature bit to turn it on. Then vhost may benefit from this. > > >>> Ran performance tests with the following configuration: >>> >>> - my packet generator -> tap1 -> br0 -> tap0 -> my packet consumer >>> - sender pinned to one core and receiver pinned to another core >>> - sender send small UDP packets (64 bytes total) as fast as it can >>> - sandy bridge cores >>> - throughput are receiver side goodput numbers >>> >>> The results are >>> >>> baseline: 757k pkts/sec, cpu utilization at 1.54 cpus >>> changed: 804k pkts/sec, cpu utilization at 1.57 cpus >>> >>> The performance difference is largely determined by packet rate and >>> inter-cpu communication cost. For example, if the sender and >>> receiver are pinned to different cpu sockets, the results are >>> >>> baseline: 558k pkts/sec, cpu utilization at 1.71 cpus >>> changed: 690k pkts/sec, cpu utilization at 1.67 cpus >> So I believe your consumer is using blocking reads. How about re-test >> with non blocking reads and re-test to make sure no regression? > > I tested non blocking read and found no regression. However the sender > is the bottleneck in my case so packet blasting is not a good test for > non blocking mode. I switched to RR / ping pong type of traffic through tap. > The packet rates for both cases are ~477k and the difference is way > below noise. > > >>> Co-authored-by: Eric Dumazet >>> Signed-off-by: Xi Wang >>> --- >>> drivers/net/tun.c | 68 +++++++++++++++++++++---------------------------------- >>> 1 file changed, 26 insertions(+), 42 deletions(-) >>> >>> diff --git a/drivers/net/tun.c b/drivers/net/tun.c >>> index ee328ba..cb25385 100644 >>> --- a/drivers/net/tun.c >>> +++ b/drivers/net/tun.c >>> @@ -133,8 +133,7 @@ struct tap_filter { >>> struct tun_file { >>> struct sock sk; >>> struct socket socket; >>> - struct socket_wq wq; >>> - struct tun_struct __rcu *tun; >>> + struct tun_struct __rcu *tun ____cacheline_aligned_in_smp; >> This seems a optimization which is un-related to the topic. May send as >> another patch but did you really see improvement for this? > > There is an ~1% difference (not as reliable as other data since the difference > is small). This is not a major performance contributor. > > >>> struct net *net; >>> struct fasync_struct *fasync; >>> /* only used for fasnyc */ >>> @@ -498,12 +497,12 @@ static void tun_detach_all(struct net_device *dev) >>> for (i = 0; i < n; i++) { >>> tfile = rtnl_dereference(tun->tfiles[i]); >>> BUG_ON(!tfile); >>> - wake_up_all(&tfile->wq.wait); >>> + tfile->socket.sk->sk_data_ready(tfile->socket.sk); >>> RCU_INIT_POINTER(tfile->tun, NULL); >>> --tun->numqueues; >>> } >>> list_for_each_entry(tfile, &tun->disabled, next) { >>> - wake_up_all(&tfile->wq.wait); >>> + tfile->socket.sk->sk_data_ready(tfile->socket.sk); >>> RCU_INIT_POINTER(tfile->tun, NULL); >>> } >>> BUG_ON(tun->numqueues != 0); >>> @@ -807,8 +806,7 @@ static netdev_tx_t tun_net_xmit(struct sk_buff *skb, struct net_device *dev) >>> /* Notify and wake up reader process */ >>> if (tfile->flags & TUN_FASYNC) >>> kill_fasync(&tfile->fasync, SIGIO, POLL_IN); >>> - wake_up_interruptible_poll(&tfile->wq.wait, POLLIN | >>> - POLLRDNORM | POLLRDBAND); >>> + tfile->socket.sk->sk_data_ready(tfile->socket.sk); >>> >>> rcu_read_unlock(); >>> return NETDEV_TX_OK; >>> @@ -965,7 +963,7 @@ static unsigned int tun_chr_poll(struct file *file, poll_table *wait) >>> >>> tun_debug(KERN_INFO, tun, "tun_chr_poll\n"); >>> >>> - poll_wait(file, &tfile->wq.wait, wait); >>> + poll_wait(file, sk_sleep(sk), wait); >>> >>> if (!skb_queue_empty(&sk->sk_receive_queue)) >>> mask |= POLLIN | POLLRDNORM; >>> @@ -1330,46 +1328,21 @@ done: >>> static ssize_t tun_do_read(struct tun_struct *tun, struct tun_file *tfile, >>> const struct iovec *iv, ssize_t len, int noblock) >>> { >>> - DECLARE_WAITQUEUE(wait, current); >>> struct sk_buff *skb; >>> ssize_t ret = 0; >>> + int peeked, err, off = 0; >>> >>> tun_debug(KERN_INFO, tun, "tun_do_read\n"); >>> >>> - if (unlikely(!noblock)) >>> - add_wait_queue(&tfile->wq.wait, &wait); >>> - while (len) { >>> - if (unlikely(!noblock)) >>> - current->state = TASK_INTERRUPTIBLE; >>> - >>> - /* Read frames from the queue */ >>> - if (!(skb = skb_dequeue(&tfile->socket.sk->sk_receive_queue))) { >>> - if (noblock) { >>> - ret = -EAGAIN; >>> - break; >>> - } >>> - if (signal_pending(current)) { >>> - ret = -ERESTARTSYS; >>> - break; >>> - } >>> - if (tun->dev->reg_state != NETREG_REGISTERED) { >>> - ret = -EIO; >>> - break; >>> - } >>> - >>> - /* Nothing to read, let's sleep */ >>> - schedule(); >>> - continue; >>> - } >>> + if (!len) >>> + return ret; >>> >>> + /* Read frames from queue */ >>> + skb = __skb_recv_datagram(tfile->socket.sk, noblock ? MSG_DONTWAIT : 0, >>> + &peeked, &off, &err); >>> + if (skb) { >> This changes the userspace ABI a little bit. Originally, userspace can >> see different error codes and do responds, but here it can only see zero. > > Thanks for catching this! Seems forwarding the &err parameter of > __skb_recv_datagram > should get the most of the error code compatibility back? Seems not, -ERESTARTSYS and EIO were missed. > I'll check > related code. > > >>> ret = tun_put_user(tun, tfile, skb, iv, len); >>> kfree_skb(skb); >>> - break; >>> - } >>> - >>> - if (unlikely(!noblock)) { >>> - current->state = TASK_RUNNING; >>> - remove_wait_queue(&tfile->wq.wait, &wait); >>> } >>> >>> return ret; >>> @@ -2187,20 +2160,28 @@ out: >>> static int tun_chr_open(struct inode *inode, struct file * file) >>> { >>> struct tun_file *tfile; >>> + struct socket_wq *wq; >>> >>> DBG1(KERN_INFO, "tunX: tun_chr_open\n"); >>> >>> + wq = kzalloc(sizeof(*wq), GFP_KERNEL); >>> + if (!wq) >>> + return -ENOMEM; >>> + >> Why not just reusing the socket_wq structure inside tun_file structure >> like what we did in the past? > > There is no strong reason for going either way. Changing to dynamic allocation > is based on: Less chance of cacheline contention and syncing the code pattern > with core stack. It's seems another possible optimization un-related to the topic, better send with another patch. But I suspect how much it will help for the performance. Checking the other socket implementation such as af unix socket, the socket_wq structure were also embedded in the parent socket structure. > > > -Xi > -- > To unsubscribe from this list: send the line "unsubscribe netdev" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html