netdev.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Jesper Dangaard Brouer <brouer@redhat.com>
To: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Jason Wang <jasowang@redhat.com>,
	Eric Dumazet <eric.dumazet@gmail.com>,
	davem@davemloft.net, netdev@vger.kernel.org,
	linux-kernel@vger.kernel.org,
	Steven Rostedt <rostedt@goodmis.org>,
	brouer@redhat.com
Subject: Re: [PATCH net-next] tuntap: introduce tx skb ring
Date: Thu, 19 May 2016 13:59:30 +0200	[thread overview]
Message-ID: <20160519135930.2a47163d@redhat.com> (raw)
In-Reply-To: <20160518191148-mutt-send-email-mst@redhat.com>

On Wed, 18 May 2016 19:26:38 +0300
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Wed, May 18, 2016 at 10:13:59AM +0200, Jesper Dangaard Brouer wrote:
> > I agree. It is sad to see everybody is implementing the same thing,
> > open coding an array/circular based ring buffer.  This kind of code is
> > hard to maintain and get right with barriers etc.  We can achieve the
> > same performance with a generic implementation, by inlining the help
> > function calls.  
> 
> So my testing seems to show that at least for the common usecase
> in networking, which isn't lockless, circular buffer
> with indices does not perform that well, because
> each index access causes a cache line to bounce between
> CPUs, and index access causes stalls due to the dependency.

Yes, I have also noticed this.

For example my qmempool implementation which is based on alf_queue,
does not scale perfectly, likely because of this.

Concurrency benchmark:
 https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/mm/qmempool_bench_parallel.c

for N in $(seq 1 8); do modprobe qmempool_bench_parallel parallel_cpus=$N run_flags=$((2#1000)) ; rmmod qmempool_bench_parallel && dmesg | tail -n 6 | grep parallel_qmempool_pattern_softirq_inline | awk '{print "Qmempool parallel "$8," ",$5," ",$6," ",$7}'; done 

Qmempool parallel CPUs:1   Average:   25   cycles(tsc)
Qmempool parallel CPUs:2   Average:   54   cycles(tsc)
Qmempool parallel CPUs:3   Average:   68   cycles(tsc)
Qmempool parallel CPUs:4   Average:   98   cycles(tsc)
Qmempool parallel CPUs:5   Average:   112   cycles(tsc)
Qmempool parallel CPUs:6   Average:   136   cycles(tsc)
Qmempool parallel CPUs:7   Average:   168   cycles(tsc)
Qmempool parallel CPUs:8   Average:   222   cycles(tsc)

The test above  does 1024 allocs followed by 1024 frees, to a qmempool,
which will cache 64 objects locally before accessing the shared
alf_queue pool (func run_bench_N_pattern_qmempool).


> By comparison, an array of pointers where NULL means invalid
> and !NULL means valid, can be updated without messing up barriers
> at all and does not have this issue.

We should verify this by benchmarking.  Once you have fixed the bug
Eric pointed out, I can try to benchmark this for you...


> You also mentioned cache pressure caused by using large queues, and I
> think it's a significant issue. tun has a queue of 1000 entries by
> default and that's 8K already.
> 
> So, I had an idea: with an array of pointers we could actually use
> only part of the ring as long as it stays mostly empty.
> We do want to fill at least two cache lines to prevent producer
> and consumer from writing over the same cache line all the time.
> This is SKB_ARRAY_MIN_SIZE logic below.

I really like this idea.  The only problem is that performance
characteristics will change according to load, which makes it harder to
benchmark, and verify that both situations are covered.  I guess, in a
micro-benchmark we could make sure that be cover both cases.  In
real-life scenarios it might be harder...


> Pls take a look at the implementation below.  It's a straight port from virtio
> unit test, so should work fine, except the SKB_ARRAY_MIN_SIZE hack that
> I added.  Today I run out of time for testing this.  Posting for early
> flames/feedback.
> 
> It's using skb pointers but we switching to void * would be easy at cost
> of type safety, though it appears that people want lockless  push
> etc so I'm not sure of the value.
> 
> --->  
> skb_array: array based FIFO for skbs
> 
> A simple array based FIFO of pointers.
> Intended for net stack so uses skbs for type
> safety, but we can replace with with void *
> if others find it useful outside of net stack.
> 
> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> 
> ---
> 
> diff --git a/include/linux/skb_array.h b/include/linux/skb_array.h
> new file mode 100644
> index 0000000..a67cc8b
> --- /dev/null
> +++ b/include/linux/skb_array.h
> @@ -0,0 +1,116 @@
> +/*
> + * See Documentation/skbular-buffers.txt for more information.
> + */
> +
> +#ifndef _LINUX_SKB_ARRAY_H
> +#define _LINUX_SKB_ARRAY_H 1
> +
> +#include <linux/spinlock.h>
> +#include <linux/cache.h>
> +#include <linux/types.h>
> +#include <linux/compiler.h>
> +#include <linux/cache.h>
> +#include <linux/slab.h>
> +#include <asm/errno.h>
> +
> +struct sk_buff;
> +
> +struct skb_array {
> +	int producer ____cacheline_aligned_in_smp;
> +	spinlock_t producer_lock;
> +	int consumer ____cacheline_aligned_in_smp;
> +	spinlock_t consumer_lock;
> +	/* Shared consumer/producer data */
> +	int size ____cacheline_aligned_in_smp; /* max entries in queue */
> +	struct sk_buff **queue;
> +};
> +
> +#define SKB_ARRAY_MIN_SIZE (2 * (0x1 << cache_line_size()) / \
> +			    sizeof (struct sk_buff *))
> +
> +static inline int __skb_array_produce(struct skb_array *a,
> +				       struct sk_buff *skb)
> +{
> +	/* Try to start from beginning: good for cache utilization as we'll
> +	 * keep reusing the same cache line.
> +	 * Produce at least SKB_ARRAY_MIN_SIZE entries before trying to do this,
> +	 * to reduce bouncing cache lines between them.
> +	 */
> +	if (a->producer >= SKB_ARRAY_MIN_SIZE && !a->queue[0])
> +		a->producer = 0;
> +	if (a->queue[a->producer])
> +		return -ENOSPC;
> +	a->queue[a->producer] = skb;
> +	if (unlikely(++a->producer > a->size))
> +		a->producer = 0;
> +	return 0;
> +}
> +
> +static inline int skb_array_produce_bh(struct skb_array *a,
> +				       struct sk_buff *skb)
> +{
> +	int ret;
> +
> +	spin_lock_bh(&a->producer_lock);
> +	ret = __skb_array_produce(a, skb);
> +	spin_unlock_bh(&a->producer_lock);
> +
> +	return ret;
> +}
> +
> +static inline struct sk_buff *__skb_array_peek(struct skb_array *a)
> +{
> +	if (a->queue[a->consumer])
> +		return a->queue[a->consumer];
> +
> +	/* Check whether producer started at the beginning. */
> +	if (unlikely(a->consumer >= SKB_ARRAY_MIN_SIZE && a->queue[0])) {
> +		a->consumer = 0;
> +		return a->queue[0];
> +	}
> +
> +	return NULL;
> +}
> +
> +static inline void __skb_array_consume(struct skb_array *a)
> +{
> +	a->queue[a->consumer++] = NULL;
> +	if (unlikely(++a->consumer > a->size))
> +		a->consumer = 0;
> +}
> +
> +static inline struct sk_buff *skb_array_consume_bh(struct skb_array *a)
> +{
> +	struct sk_buff *skb;
> +
> +	spin_lock_bh(&a->producer_lock);
> +	skb = __skb_array_peek(a);
> +	if (skb)
> +		__skb_array_consume(a);
> +	spin_unlock_bh(&a->producer_lock);
> +
> +	return skb;
> +}
> +
> +static inline int skb_array_init(struct skb_array *a, int size, gfp_t gfp)
> +{
> +	a->queue = kmalloc(ALIGN(size * sizeof *(a->queue), SMP_CACHE_BYTES),
> +			   gfp);
> +	if (!a->queue)
> +		return -ENOMEM;
> +
> +	a->size = size;
> +	a->producer = a->consumer = 0;
> +	spin_lock_init(&a->producer_lock);
> +	spin_lock_init(&a->consumer_lock);
> +
> +	return 0;
> +}
> +
> +static inline void skb_array_cleanup(struct skb_array *a)
> +{
> +	kfree(a->queue);
> +}
> +
> +#endif /* _LINUX_SKB_ARRAY_H  */



-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  Author of http://www.iptv-analyzer.org
  LinkedIn: http://www.linkedin.com/in/brouer

  parent reply	other threads:[~2016-05-19 11:59 UTC|newest]

Thread overview: 21+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2016-05-16  1:17 [PATCH net-next] tuntap: introduce tx skb ring Jason Wang
2016-05-16  3:56 ` Eric Dumazet
2016-05-16  7:51   ` Jason Wang
2016-05-18  8:13     ` Jesper Dangaard Brouer
2016-05-18  8:23       ` Michael S. Tsirkin
2016-05-18 10:23       ` Jason Wang
2016-05-18 11:52         ` Steven Rostedt
2016-05-18 16:26       ` Michael S. Tsirkin
2016-05-18 16:41         ` Eric Dumazet
2016-05-18 16:46           ` Michael S. Tsirkin
2016-05-19 11:59         ` Jesper Dangaard Brouer [this message]
2016-05-16  4:23 ` Michael S. Tsirkin
2016-05-16  7:52   ` Jason Wang
2016-05-16  8:08     ` Michael S. Tsirkin
2016-05-17  1:38       ` Jason Wang
2016-05-18  8:16         ` Jesper Dangaard Brouer
2016-05-18  8:21           ` Michael S. Tsirkin
2016-05-18  9:21             ` Jesper Dangaard Brouer
2016-05-18  9:55               ` Michael S. Tsirkin
2016-05-18 10:42                 ` Jason Wang
2016-05-18 10:58                   ` Michael S. Tsirkin

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20160519135930.2a47163d@redhat.com \
    --to=brouer@redhat.com \
    --cc=davem@davemloft.net \
    --cc=eric.dumazet@gmail.com \
    --cc=jasowang@redhat.com \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mst@redhat.com \
    --cc=netdev@vger.kernel.org \
    --cc=rostedt@goodmis.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).