Network performance regression with vde_switch

I've noticed a significant network performance regression when using vde_switch, starting about one week ago (10/05/2012); before that date, I used to get about 1.5 Gbits host to guest, but now I can only get about 320 Mbits; I didn't find any modification in net/vde.*, just in hw/virtio*.

My command line: 
 qemu-system-i386 -cdrom /bpd/bpd.iso -m 512 -boot d -enable-kvm \
  -localtime -ctrl-grab -usbdevice tablet \
  -device virtio-net-pci,mac=52:54:00:18:01:01,netdev=vde0,tx=bh,ioeventfd=on,x-txburst=32 \
  -netdev vde,id=vde0 -vga std -tb-size 2M -cpu host -clock unix

My host runs a kernel 3.6.1 and my guest runs a kernel 3.5.4; the same problem happens with other host and guest versions, too.

I know there are better ways of running a guest, but using vde I get a cleaner environment in the host (just one tun/tap interface to manage...), which is quite good when running some accademic experiments.

Interestingly, at the same time I've noticed a performance enhancement of about 25~30 % when using a tun/tap interface, bridged or not.

Thank you, very much.

Edivaldo de Araujo Pereira

On Fri, Oct 12, 2012 at 05:34:23PM -0000, Edivaldo de Araujo Pereira wrote:
> I've noticed a significant network performance regression when using
> vde_switch, starting about one week ago (10/05/2012); before that date,
> I used to get about 1.5 Gbits host to guest, but now I can only get
> about 320 Mbits; I didn't find any modification in net/vde.*, just in
> hw/virtio*.
> 
> My command line: 
>  qemu-system-i386 -cdrom /bpd/bpd.iso -m 512 -boot d -enable-kvm \
>   -localtime -ctrl-grab -usbdevice tablet \
>   -device virtio-net-pci,mac=52:54:00:18:01:01,netdev=vde0,tx=bh,ioeventfd=on,x-txburst=32 \
>   -netdev vde,id=vde0 -vga std -tb-size 2M -cpu host -clock unix
> 
> My host runs a kernel 3.6.1 and my guest runs a kernel 3.5.4; the same
> problem happens with other host and guest versions, too.
> 
> I know there are better ways of running a guest, but using vde I get a
> cleaner environment in the host (just one tun/tap interface to
> manage...), which is quite good when running some accademic experiments.
> 
> Interestingly, at the same time I've noticed a performance enhancement
> of about 25~30 % when using a tun/tap interface, bridged or not.

Hi Edivaldo,
It would be great if you can help find the commit that caused this
regression.

The basic process is:

1. Identify a QEMU release or git tree that gives you 1.5 Gbit/s.
2. Double-check that qemu.git/master suffers reduced performance.
3. git bisect start <bad> <good>
   where <bad> and <good> are the git commits that show differing
   performance (for example, bad=HEAD good=v1.1.0)

Then git will step through the commit history and ask you to test at
each step.  (This is a binary search so even finding regressions that
happened many commits ago requires few steps.)

You can read more about git-bisect(1) here:
http://git-scm.com/book/en/Git-Tools-Debugging-with-Git#Binary-Search
http://www.kernel.org/pub/software/scm/git/docs/git-bisect.html

The end result is the commit introduced the regression.  Please post
what you find!

Stefan


Hi Stefan,

Thank you, very much for taking the time to help me, and excuse me for not seeing your answer early... 

I've run the procedure you pointed me out, and the result is:

0d8d7690850eb0cf2b2b60933cf47669a6b6f18f is the first bad commit
commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
Author: Amit Shah <email address hidden>
Date:   Tue Sep 25 00:05:15 2012 +0530

    virtio: Introduce virtqueue_get_avail_bytes()

    The current virtqueue_avail_bytes() is oddly named, and checks if a
    particular number of bytes are available in a vq.  A better API is to
    fetch the number of bytes available in the vq, and let the caller do
    what's interesting with the numbers.

    Introduce virtqueue_get_avail_bytes(), which returns the number of bytes
    for buffers marked for both, in as well as out.  virtqueue_avail_bytes()
    is made a wrapper over this new function.

    Signed-off-by: Amit Shah <email address hidden>
    Signed-off-by: Michael S. Tsirkin <email address hidden>

:040000 040000 1a58b06a228651cf844621d9ee2f49b525e36c93 e09ea66ce7f6874921670b6aeab5bea921a5227d M      hw

I tried to revert that patch in the latest version, but it obviously didnt work; I'm trying to figure out the problem, but I don't know very well the souce code, so I think it's going to take some time. For now, it's all I could do.

Thank you, again.
Edivaldo



On Mon, Oct 15, 2012 at 09:46:06PM -0000, Edivaldo de Araujo Pereira wrote:
> Hi Stefan,
> 
> Thank you, very much for taking the time to help me, and excuse me for
> not seeing your answer early...
> 
> I've run the procedure you pointed me out, and the result is:
> 
> 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f is the first bad commit
> commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
> Author: Amit Shah <email address hidden>
> Date:   Tue Sep 25 00:05:15 2012 +0530
> 
>     virtio: Introduce virtqueue_get_avail_bytes()
> 
>     The current virtqueue_avail_bytes() is oddly named, and checks if a
>     particular number of bytes are available in a vq.  A better API is to
>     fetch the number of bytes available in the vq, and let the caller do
>     what's interesting with the numbers.
> 
>     Introduce virtqueue_get_avail_bytes(), which returns the number of bytes
>     for buffers marked for both, in as well as out.  virtqueue_avail_bytes()
>     is made a wrapper over this new function.
> 
>     Signed-off-by: Amit Shah <email address hidden>
>     Signed-off-by: Michael S. Tsirkin <email address hidden>
> 
> :040000 040000 1a58b06a228651cf844621d9ee2f49b525e36c93
> e09ea66ce7f6874921670b6aeab5bea921a5227d M      hw
> 
> I tried to revert that patch in the latest version, but it obviously
> didnt work; I'm trying to figure out the problem, but I don't know very
> well the souce code, so I think it's going to take some time. For now,
> it's all I could do.

After git-bisect(1) completes it is good to sanity-check the result by
manually testing 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f^ (the commit
just before the bad commit) and 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
(the bad commit).

This will verify that the commit indeed introduces the regression.  I
suggest doing this just to be sure that you've found the bad commit.

Regarding this commit, I notice two things:

1. We will now loop over all vring descriptors because we calculate the
   total in/out length instead of returning early as soon as we see
   there is enough space.  Maybe this makes a difference, although I'm a
   little surprised you see such a huge regression.

2. The comparision semantics have changed from:

     (in_total += vring_desc_len(desc_pa, i)) >= in_bytes

   to:

     (in_bytes && in_bytes < in_total)

   Notice that virtqueue_avail_bytes() now returns 0 when in_bytes ==
   in_total.  Previously, it would return 1.  Perhaps we are starving or
   delaying I/O due to this comparison change.  You can easily change
   '<' to '<=' to see if it fixes the issue.

Stefan


Hi Stefan

I finally could revert that commit in the latest snapshot; problem was I needed to revert one later, that modified hw/virtio-serial-bus.c accordingly; after that reversion, the regression in network performance went completely away; this confirms my previous identification of the commit that caused it.

Additionally, I tested your last suggestion, to change '<' to '<=', and that didn't help; the problem was still there.

By the way, the performance gain I observed ,of about 25~30 % when using a tun/tap, was in fact just apparent, because it was result of a greater use of cpu, so it was achieved only when the host was idle; when I stress the host, with a lot of concurrent guests generating traffic, there is no gain at all.

Just for confirmation, this is the performance I get with latest snapshot (8b4a3df8081f3e6f1061ed5cbb303ad623ade66b) running wget in the guest:

$ wget -O /dev/null http://172.18.1.1/bpd.iso
--2012-10-16 09:10:18--  http://172.18.1.1/bpd.iso
Connecting to 172.18.1.1:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 358979584 (342M) [application/x-iso9660-image]
Saving to: `/dev/null'
100%[======================================================>] 358.979.584 29,7M/s   in 11s
2012-10-16 09:10:29 (30,3 MB/s) - `/dev/null' saved [358979584/358979584]

The same wget, using the same snapshot, but with that commit reverted is:

$ wget -O /dev/null http://172.18.1.1/bpd.iso
--2012-10-16 09:15:12--  http://172.18.1.1/bpd.iso
Connecting to 172.18.1.1:80... connected.
HTTP request sent, awaiting response... 200 OK
Length: 358979584 (342M) [application/x-iso9660-image]
Saving to: `/dev/null'
100%[======================================================>] 358.979.584  180M/s   in 1,9s
2012-10-16 09:15:14 (180 MB/s) - `/dev/null' saved [358979584/358979584]

So, as I can see, there is no doubt: that commit is the culprit; as it was intended to be related just to the quality of the source code (at least as I can see), but implied in such a cost in performance, I think it would be better to revert it.

Thank you very much, again.
Edivaldo

On Tue, Oct 16, 2012 at 2:23 PM, Edivaldo de Araujo Pereira
<email address hidden> wrote:
> Hi Stefan
>
> I finally could revert that commit in the latest snapshot; problem was I
> needed to revert one later, that modified hw/virtio-serial-bus.c
> accordingly; after that reversion, the regression in network performance
> went completely away; this confirms my previous identification of the
> commit that caused it.
>
> Additionally, I tested your last suggestion, to change '<' to '<=', and
> that didn't help; the problem was still there.
>
> By the way, the performance gain I observed ,of about 25~30 % when using
> a tun/tap, was in fact just apparent, because it was result of a greater
> use of cpu, so it was achieved only when the host was idle; when I
> stress the host, with a lot of concurrent guests generating traffic,
> there is no gain at all.
>
> Just for confirmation, this is the performance I get with latest
> snapshot (8b4a3df8081f3e6f1061ed5cbb303ad623ade66b) running wget in the
> guest:
>
> $ wget -O /dev/null http://172.18.1.1/bpd.iso
> --2012-10-16 09:10:18--  http://172.18.1.1/bpd.iso
> Connecting to 172.18.1.1:80... connected.
> HTTP request sent, awaiting response... 200 OK
> Length: 358979584 (342M) [application/x-iso9660-image]
> Saving to: `/dev/null'
> 100%[======================================================>] 358.979.584 29,7M/s   in 11s
> 2012-10-16 09:10:29 (30,3 MB/s) - `/dev/null' saved [358979584/358979584]
>
> The same wget, using the same snapshot, but with that commit reverted
> is:
>
> $ wget -O /dev/null http://172.18.1.1/bpd.iso
> --2012-10-16 09:15:12--  http://172.18.1.1/bpd.iso
> Connecting to 172.18.1.1:80... connected.
> HTTP request sent, awaiting response... 200 OK
> Length: 358979584 (342M) [application/x-iso9660-image]
> Saving to: `/dev/null'
> 100%[======================================================>] 358.979.584  180M/s   in 1,9s
> 2012-10-16 09:15:14 (180 MB/s) - `/dev/null' saved [358979584/358979584]
>
> So, as I can see, there is no doubt: that commit is the culprit; as it
> was intended to be related just to the quality of the source code (at
> least as I can see), but implied in such a cost in performance, I think
> it would be better to revert it.

Hi Amit,
Edivaldo has identified the following commit responsible for a network
performance regression he sees:

commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
Author: Amit Shah <email address hidden>
Date:   Tue Sep 25 00:05:15 2012 +0530

    virtio: Introduce virtqueue_get_avail_bytes()

I guess this is because we now iterate the entire descriptor chain to
check available space instead of returning early.

Do you want to propose a patch to fix it?

Stefan


Dear Amit,

On a suggestion of Stefan, I've already tested the modification in you patch, and it didn't work; but for confirmation I tested it once again, on the latest snapshot; same result, that is, it didn't work; the problem is still there.

I didn't take enough time to uderstand the code, so unfortunately I fear there is not much I could do to solve the problem, apart from trying your suggestions. But I'll try to spend a little more time on it, until we find a solution.

Thank you very much.

Edivaldo

--- Em seg, 22/10/12, Amit Shah <email address hidden> escreveu:

> De: Amit Shah <email address hidden>
> Assunto: Re: [Qemu-devel] [Bug 1066055] Re: Network performance regression with vde_switch
> Para: "Stefan Hajnoczi" <email address hidden>
> Cc: "Bug 1066055" <email address hidden>, <email address hidden>, <email address hidden>
> Data: Segunda-feira, 22 de Outubro de 2012, 4:18
> On (Tue) 16 Oct 2012 [09:48:09],
> Stefan Hajnoczi wrote:
> > On Mon, Oct 15, 2012 at 09:46:06PM -0000, Edivaldo de
> Araujo Pereira wrote:
> > > Hi Stefan,
> > > 
> > > Thank you, very much for taking the time to help
> me, and excuse me for
> > > not seeing your answer early...
> > > 
> > > I've run the procedure you pointed me out, and the
> result is:
> > > 
> > > 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f is the
> first bad commit
> > > commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
> > > Author: Amit Shah <email address hidden>
> > > Date:   Tue Sep 25 00:05:15 2012
> +0530
> > > 
> > >     virtio: Introduce
> virtqueue_get_avail_bytes()
> > > 
> > >     The current
> virtqueue_avail_bytes() is oddly named, and checks if a
> > >     particular number of bytes
> are available in a vq.  A better API is to
> > >     fetch the number of bytes
> available in the vq, and let the caller do
> > >     what's interesting with
> the numbers.
> > > 
> > >     Introduce
> virtqueue_get_avail_bytes(), which returns the number of
> bytes
> > >     for buffers marked for
> both, in as well as out.  virtqueue_avail_bytes()
> > >     is made a wrapper over
> this new function.
> > > 
> > >     Signed-off-by: Amit Shah
> <email address hidden>
> > >     Signed-off-by: Michael S.
> Tsirkin <email address hidden>
> > > 
> > > :040000 040000
> 1a58b06a228651cf844621d9ee2f49b525e36c93
> > > e09ea66ce7f6874921670b6aeab5bea921a5227d M 
>     hw
> > > 
> > > I tried to revert that patch in the latest
> version, but it obviously
> > > didnt work; I'm trying to figure out the problem,
> but I don't know very
> > > well the souce code, so I think it's going to take
> some time. For now,
> > > it's all I could do.
> > 
> > After git-bisect(1) completes it is good to
> sanity-check the result by
> > manually testing
> 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f^ (the commit
> > just before the bad commit) and
> 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
> > (the bad commit).
> > 
> > This will verify that the commit indeed introduces the
> regression.  I
> > suggest doing this just to be sure that you've found
> the bad commit.
> > 
> > Regarding this commit, I notice two things:
> > 
> > 1. We will now loop over all vring descriptors because
> we calculate the
> >    total in/out length instead of returning
> early as soon as we see
> >    there is enough space.  Maybe this
> makes a difference, although I'm a
> >    little surprised you see such a huge
> regression.
> > 
> > 2. The comparision semantics have changed from:
> > 
> >      (in_total +=
> vring_desc_len(desc_pa, i)) >= in_bytes
> > 
> >    to:
> > 
> >      (in_bytes && in_bytes <
> in_total)
> > 
> >    Notice that virtqueue_avail_bytes() now
> returns 0 when in_bytes ==
> >    in_total.  Previously, it would
> return 1.  Perhaps we are starving or
> >    delaying I/O due to this comparison
> change.  You can easily change
> >    '<' to '<=' to see if it fixes the
> issue.
> 
> Hi Edivaldo,
> 
> Can you try the following patch, that will confirm if it's
> the
> descriptor walk or the botched compare that's causing the
> regression.
> 
> Thanks,
> 
> diff --git a/hw/virtio.c b/hw/virtio.c
> index 6821092..bb08ed8 100644
> --- a/hw/virtio.c
> +++ b/hw/virtio.c
> @@ -406,8 +406,8 @@ int virtqueue_avail_bytes(VirtQueue *vq,
> unsigned int in_bytes,
>      unsigned int in_total, out_total;
>  
>      virtqueue_get_avail_bytes(vq,
> &in_total, &out_total);
> -    if ((in_bytes && in_bytes <
> in_total)
> -        || (out_bytes &&
> out_bytes < out_total)) {
> +    if ((in_bytes && in_bytes <=
> in_total)
> +        || (out_bytes &&
> out_bytes <= out_total)) {
>          return 1;
>      }
>      return 0;
> 
> 
>         Amit
> 


On Mon, Oct 22, 2012 at 06:50:00AM -0700, Edivaldo de Araujo Pereira wrote:
> I didn't take enough time to uderstand the code, so unfortunately I fear there is not much I could do to solve the problem, apart from trying your suggestions. But I'll try to spend a little more time on it, until we find a solution.

I've thought a little about how to approach this.  Amit, here's a brain
dump:

The simplest solution is to make virtqueue_avail_bytes() use the old
behavior of stopping early.

However, I wonder if we can actually *improve* performance of existing
code by changing virtio-net.c:virtio_net_receive().  The intuition is
that calling virtio_net_has_buffers() (internally calls
virtqueue_avail_bytes()) followed by virtqueue_pop() is suboptimal
because we're repeatedly traversing the descriptor chain.

We can get rid of this repetition.  A side-effect of this is that we no
longer need to call virtqueue_avail_bytes() from virtio-net.c.  Here's
how:

The common case in virtio_net_receive() is that we have buffers and they
are large enough for the received packet.  So to optimize for this case:

1. Take the VirtQueueElement off the vring but don't increment
   last_avail_idx yet.  (This is essentially a "peek" operation.)

2. If there is an error or we drop the packet because the
   VirtQueueElement is too small, just bail out and we'll grab the same
   VirtQueueElement again next time.

3. When we've committed filling in this VirtQueueElement, increment
   last_avail_idx.  This is the point of no return.

Essentially we're splitting pop() into peek() and consume().  Peek()
grabs the VirtQueueElement but does not increment last_avail_idx.
Consume() simply increments last_avail_idx and maybe the EVENT_IDX
optimization stuff.

Whether this will improve performance, I'm not sure.  Perhaps
virtio_net_has_buffers() pulls most descriptors into the CPU's cache and
following up with virtqueue_pop() is very cheap already.  But the idea
here is to avoid the virtio_net_has_buffers() because we'll find out
soon enough when we try to pop :).

Another approach would be to drop virtio_net_has_buffers() but continue
to use virtqueue_pop().  We'd keep the same VirtQueueElem stashed in
VirtIONet across virtio_net_receive() calls in the case where we drop
the packet.  I don't like this approach very much though because it gets
tricky when the guest modifies the vring memory, resets the virtio
device, etc across calls.

Stefan


On Thu, Nov 1, 2012 at 5:07 PM, Michael S. Tsirkin <email address hidden> wrote:
> Commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f introduced
> a regression in virtio-net performance because it looks
> into the ring aggressively while we really only care
> about a single packet worth of buffers.
> To fix, add parameters limiting lookahead, and
> use in virtqueue_avail_bytes.
>
> Signed-off-by: Michael S. Tsirkin <email address hidden>
> Reported-by: Edivaldo de Araujo Pereira <email address hidden>

Nice, much simpler than the ideas I had.

Reviewed-by: Stefan Hajnoczi <email address hidden>


On Fri, Nov 2, 2012 at 3:48 PM, Michael S. Tsirkin <email address hidden> wrote:
> On Fri, Nov 02, 2012 at 11:18:18AM +0100, Stefan Hajnoczi wrote:
>> On Thu, Nov 1, 2012 at 5:07 PM, Michael S. Tsirkin <email address hidden> wrote:
>> > Commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f introduced
>> > a regression in virtio-net performance because it looks
>> > into the ring aggressively while we really only care
>> > about a single packet worth of buffers.
>> > To fix, add parameters limiting lookahead, and
>> > use in virtqueue_avail_bytes.
>> >
>> > Signed-off-by: Michael S. Tsirkin <email address hidden>
>> > Reported-by: Edivaldo de Araujo Pereira <email address hidden>
>>
>> Nice, much simpler than the ideas I had.
>>
>> Reviewed-by: Stefan Hajnoczi <email address hidden>
>
> Anthony could you apply this out of band please so this stops
> biting people?

Especially for the 1.3 release so that we don't have a virtio
performance regression.

Stefan


Dear friends, 

Please excuse-me for not reporting earlier... I confirm that the patch by Michael really fixes the problem I've reported. The regression has gone away when I used it, so I think it is good to be applied.

Thanks,

Edivaldo de Araújo Pereira


--- Em ter, 27/11/12, Michael S. Tsirkin <email address hidden> escreveu:

> De: Michael S. Tsirkin <email address hidden>
> Assunto: Re: [PATCH] virtio: limit avail bytes lookahead
> Para: "Amit Shah" <email address hidden>
> Cc: "Stefan Hajnoczi" <email address hidden>, "Edivaldo de Araujo Pereira" <email address hidden>, <email address hidden>, "Anthony Liguori" <email address hidden>, "Bug 1066055" <email address hidden>
> Data: Terça-feira, 27 de Novembro de 2012, 8:25
> On Thu, Nov 01, 2012 at 06:07:21PM
> +0200, Michael S. Tsirkin wrote:
> > Commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f
> introduced
> > a regression in virtio-net performance because it
> looks
> > into the ring aggressively while we really only care
> > about a single packet worth of buffers.
> > To fix, add parameters limiting lookahead, and
> > use in virtqueue_avail_bytes.
> > 
> > Signed-off-by: Michael S. Tsirkin <email address hidden>
> > Reported-by: Edivaldo de Araujo Pereira <email address hidden>
> 
> Ping.
> Anthony - going to apply this?
> 
> 
> > ---
> > 
> > Edivaldo could you please confirm this fixes
> regression?
> > 
> > diff --git a/hw/virtio-serial-bus.c
> b/hw/virtio-serial-bus.c
> > index d20bd8b..a761cdc 100644
> > --- a/hw/virtio-serial-bus.c
> > +++ b/hw/virtio-serial-bus.c
> > @@ -297,7 +297,7 @@ size_t
> virtio_serial_guest_ready(VirtIOSerialPort *port)
> >      if (use_multiport(port->vser)
> && !port->guest_connected) {
> >          return 0;
> >      }
> > -    virtqueue_get_avail_bytes(vq,
> &bytes, NULL);
> > +    virtqueue_get_avail_bytes(vq,
> &bytes, NULL, UINT_MAX, 0);
> >      return bytes;
> >  }
> >  
> > diff --git a/hw/virtio.c b/hw/virtio.c
> > index ec8b7d8..f40a8c5 100644
> > --- a/hw/virtio.c
> > +++ b/hw/virtio.c
> > @@ -336,7 +336,8 @@ static unsigned
> virtqueue_next_desc(hwaddr desc_pa,
> >  }
> >  
> >  void virtqueue_get_avail_bytes(VirtQueue *vq,
> unsigned int *in_bytes,
> > -             
>              
>    unsigned int *out_bytes)
> > +             
>              
>    unsigned int *out_bytes,
> > +             
>              
>    unsigned max_in_bytes, unsigned
> max_out_bytes)
> >  {
> >      unsigned int idx;
> >      unsigned int total_bufs, in_total,
> out_total;
> > @@ -385,6 +386,9 @@ void
> virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int
> *in_bytes,
> >              } else
> {
> >               
>   out_total += vring_desc_len(desc_pa, i);
> >              }
> > +            if (in_total
> >= max_in_bytes && out_total >= max_out_bytes)
> {
> > +             
>   goto done;
> > +            }
> >          } while ((i =
> virtqueue_next_desc(desc_pa, i, max)) != max);
> >  
> >          if (!indirect)
> > @@ -392,6 +396,7 @@ void
> virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int
> *in_bytes,
> >          else
> >             
> total_bufs++;
> >      }
> > +done:
> >      if (in_bytes) {
> >          *in_bytes =
> in_total;
> >      }
> > @@ -405,12 +410,8 @@ int
> virtqueue_avail_bytes(VirtQueue *vq, unsigned int in_bytes,
> >  {
> >      unsigned int in_total, out_total;
> >  
> > -    virtqueue_get_avail_bytes(vq,
> &in_total, &out_total);
> > -    if ((in_bytes && in_bytes <
> in_total)
> > -        || (out_bytes &&
> out_bytes < out_total)) {
> > -        return 1;
> > -    }
> > -    return 0;
> > +    virtqueue_get_avail_bytes(vq,
> &in_total, &out_total, in_bytes, out_bytes);
> > +    return in_bytes <= in_total
> && out_bytes <= out_total;
> >  }
> >  
> >  void virtqueue_map_sg(struct iovec *sg, hwaddr
> *addr,
> > diff --git a/hw/virtio.h b/hw/virtio.h
> > index ac482be..1278328 100644
> > --- a/hw/virtio.h
> > +++ b/hw/virtio.h
> > @@ -150,7 +150,8 @@ int virtqueue_pop(VirtQueue *vq,
> VirtQueueElement *elem);
> >  int virtqueue_avail_bytes(VirtQueue *vq, unsigned
> int in_bytes,
> >               
>             unsigned int
> out_bytes);
> >  void virtqueue_get_avail_bytes(VirtQueue *vq,
> unsigned int *in_bytes,
> > -             
>              
>    unsigned int *out_bytes);
> > +             
>              
>    unsigned int *out_bytes,
> > +             
>              
>    unsigned max_in_bytes, unsigned
> max_out_bytes);
> >  
> >  void virtio_notify(VirtIODevice *vdev, VirtQueue
> *vq);
> >  
> 


"Michael S. Tsirkin" <email address hidden> writes:

> On Thu, Nov 01, 2012 at 06:07:21PM +0200, Michael S. Tsirkin wrote:
>> Commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f introduced
>> a regression in virtio-net performance because it looks
>> into the ring aggressively while we really only care
>> about a single packet worth of buffers.
>> To fix, add parameters limiting lookahead, and
>> use in virtqueue_avail_bytes.
>> 
>> Signed-off-by: Michael S. Tsirkin <email address hidden>
>> Reported-by: Edivaldo de Araujo Pereira <email address hidden>
>
> Ping.
> Anthony - going to apply this?

Yes, I've got it queued now.  In the future, please top post patches.

Regards,

Anthony Liguori

>
>
>> ---
>> 
>> Edivaldo could you please confirm this fixes regression?
>> 
>> diff --git a/hw/virtio-serial-bus.c b/hw/virtio-serial-bus.c
>> index d20bd8b..a761cdc 100644
>> --- a/hw/virtio-serial-bus.c
>> +++ b/hw/virtio-serial-bus.c
>> @@ -297,7 +297,7 @@ size_t virtio_serial_guest_ready(VirtIOSerialPort *port)
>>      if (use_multiport(port->vser) && !port->guest_connected) {
>>          return 0;
>>      }
>> -    virtqueue_get_avail_bytes(vq, &bytes, NULL);
>> +    virtqueue_get_avail_bytes(vq, &bytes, NULL, UINT_MAX, 0);
>>      return bytes;
>>  }
>>  
>> diff --git a/hw/virtio.c b/hw/virtio.c
>> index ec8b7d8..f40a8c5 100644
>> --- a/hw/virtio.c
>> +++ b/hw/virtio.c
>> @@ -336,7 +336,8 @@ static unsigned virtqueue_next_desc(hwaddr desc_pa,
>>  }
>>  
>>  void virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int *in_bytes,
>> -                               unsigned int *out_bytes)
>> +                               unsigned int *out_bytes,
>> +                               unsigned max_in_bytes, unsigned max_out_bytes)
>>  {
>>      unsigned int idx;
>>      unsigned int total_bufs, in_total, out_total;
>> @@ -385,6 +386,9 @@ void virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int *in_bytes,
>>              } else {
>>                  out_total += vring_desc_len(desc_pa, i);
>>              }
>> +            if (in_total >= max_in_bytes && out_total >= max_out_bytes) {
>> +                goto done;
>> +            }
>>          } while ((i = virtqueue_next_desc(desc_pa, i, max)) != max);
>>  
>>          if (!indirect)
>> @@ -392,6 +396,7 @@ void virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int *in_bytes,
>>          else
>>              total_bufs++;
>>      }
>> +done:
>>      if (in_bytes) {
>>          *in_bytes = in_total;
>>      }
>> @@ -405,12 +410,8 @@ int virtqueue_avail_bytes(VirtQueue *vq, unsigned int in_bytes,
>>  {
>>      unsigned int in_total, out_total;
>>  
>> -    virtqueue_get_avail_bytes(vq, &in_total, &out_total);
>> -    if ((in_bytes && in_bytes < in_total)
>> -        || (out_bytes && out_bytes < out_total)) {
>> -        return 1;
>> -    }
>> -    return 0;
>> +    virtqueue_get_avail_bytes(vq, &in_total, &out_total, in_bytes, out_bytes);
>> +    return in_bytes <= in_total && out_bytes <= out_total;
>>  }
>>  
>>  void virtqueue_map_sg(struct iovec *sg, hwaddr *addr,
>> diff --git a/hw/virtio.h b/hw/virtio.h
>> index ac482be..1278328 100644
>> --- a/hw/virtio.h
>> +++ b/hw/virtio.h
>> @@ -150,7 +150,8 @@ int virtqueue_pop(VirtQueue *vq, VirtQueueElement *elem);
>>  int virtqueue_avail_bytes(VirtQueue *vq, unsigned int in_bytes,
>>                            unsigned int out_bytes);
>>  void virtqueue_get_avail_bytes(VirtQueue *vq, unsigned int *in_bytes,
>> -                               unsigned int *out_bytes);
>> +                               unsigned int *out_bytes,
>> +                               unsigned max_in_bytes, unsigned max_out_bytes);
>>  
>>  void virtio_notify(VirtIODevice *vdev, VirtQueue *vq);
>>  


"Michael S. Tsirkin" <email address hidden> writes:

> On Thu, Nov 29, 2012 at 06:34:46PM +0530, Amit Shah wrote:
>> On (Wed) 28 Nov 2012 [23:53:08], Michael S. Tsirkin wrote:
>> > On Tue, Nov 27, 2012 at 06:25:04PM +0200, Michael S. Tsirkin wrote:
>> > > On Thu, Nov 01, 2012 at 06:07:21PM +0200, Michael S. Tsirkin wrote:
>> > > > Commit 0d8d7690850eb0cf2b2b60933cf47669a6b6f18f introduced
>> > > > a regression in virtio-net performance because it looks
>> > > > into the ring aggressively while we really only care
>> > > > about a single packet worth of buffers.
>> > > > To fix, add parameters limiting lookahead, and
>> > > > use in virtqueue_avail_bytes.
>> > > > 
>> > > > Signed-off-by: Michael S. Tsirkin <email address hidden>
>> > > > Reported-by: Edivaldo de Araujo Pereira <email address hidden>
>> > > 
>> > > Ping.
>> > > Anthony - going to apply this?
>> > 
>> > virtio rng was added since so naturally build broke.
>> > Here's a patch on top to fix it up. I never used virtio rng before so
>> > could not test at this hour, but it does fix the build.
>> > 
>> > I'll take a look at how to test it tomorrow but any
>> > info would be appreciated.
>> > Amit could you pls review?
>> 
>> Looks fine, I assume you will send a v2 of the patch to Anthony?
>> 
>> 		Amit
>
>
> Anthony volunteered to test this so there will only be v2 if he sees
> problems.

I need a Signed-off-by so why don't you just go ahead and send a v2 and
I'll test that.

Regards,

Anthony Liguori


The fix had been included here:
https://git.qemu.org/?p=qemu.git;a=commitdiff;h=e1f7b4812eab992de46c98b
... so closing this bug now.