Re: [PATCH v4] virtio-net: support the virtqueue coalescing moderation

[Heng Qi <hengqi@linux.alibaba.com>]

On Sun, Feb 19, 2023 at 10:06:51AM -0500, Michael S. Tsirkin wrote:
> On Sun, Feb 19, 2023 at 10:45:07PM +0800, Heng Qi wrote:
> > Currently coalescing parameters are grouped for all transmit and receive
> > virtqueues. This patch supports setting or getting the parameters for a
> > specified virtqueue, and a typical application of this function is netdim[1].
> > 
> > When the traffic between virtqueues is unbalanced, for example, one virtqueue
> > is busy and another virtqueue is idle, then it will be very useful to
> > control coalescing parameters at the virtqueue granularity.
> > 
> > [1] https://docs.kernel.org/networking/net_dim.html
> > 
> > Signed-off-by: Heng Qi <hengqi@linux.alibaba.com>
> > Reviewed-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
> > ---
> > This patch is on top of Alvaro's latest v7 patch: https://lists.oasis-open.org/archives/virtio-dev/202302/msg00431.html .
> > 
> > v3->v4:
> >     1. Include virtio_net_ctrl_coal in the virtio_net_ctrl_coal_vq structure. @Alvaro Karsz
> >     2. Add consideration of vq reset. @Michael S. Tsirkin, @Parav Pandit, @Alvaro Karsz
> >     3. Avoid too many examples by giving a comprehensive example. @Michael S. Tsirkin
> >     4. Fix typos and streamline clarifications. @Michael S. Tsirkin, @Parav Pandit, @Alvaro Karsz
> > 
> > v2->v3:
> >     1. Add the netdim link. @Parav Pandit
> >     2. VIRTIO_NET_F_VQ_NOTF_COAL no longer depends on VIRTIO_NET_F_NOTF_COAL. @Michael S. Tsirkin, @Alvaro Karsz
> >     3. _VQ_GET is explained more. @Michael S. Tsirkin
> >     4. Add more examples to avoid misunderstandings. @Michael S. Tsirkin
> >     5. Clarify some statements. @Michael S. Tsirkin, @Parav Pandit, @Alvaro Karsz
> >     6. Adjust the virtio_net_ctrl_coal_vq structure. @Michael S. Tsirkin
> >     7. Fix some typos. @Michael S. Tsirkin
> > 
> > v1->v2:
> >     1. Rename VIRTIO_NET_F_PERQUEUE_NOTF_COAL to VIRTIO_NET_F_VQ_NOTF_COAL. @Michael S. Tsirkin
> >     2. Use the \field{vqn} instead of the qid. @Michael S. Tsirkin
> >     3. Unify tx and rx control structres into one structure virtio_net_ctrl_coal_vq. @Michael S. Tsirkin
> >     4. Add a new control command VIRTIO_NET_CTRL_NOTF_COAL_VQ. @Michael S. Tsirkin, @Parav Pandit, @Alvaro Karsz
> >     5. The special value 0xFFF is removed because VIRTIO_NET_CTRL_NOTF_COAL can be used. @Alvaro Karsz
> >     6. Clarify some special scenarios. @Michael S. Tsirkin, @Parav Pandit, @Alvaro Karsz
> > 
> >  device-types/net/description.tex | 64 ++++++++++++++++++++++++++++++--
> >  1 file changed, 60 insertions(+), 4 deletions(-)
> > 
> > diff --git a/device-types/net/description.tex b/device-types/net/description.tex
> > index e71e33b..fa8bd26 100644
> > --- a/device-types/net/description.tex
> > +++ b/device-types/net/description.tex
> > @@ -83,6 +83,8 @@ \subsection{Feature bits}\label{sec:Device Types / Network Device / Feature bits
> >  \item[VIRTIO_NET_F_CTRL_MAC_ADDR(23)] Set MAC address through control
> >      channel.
> >  
> > +\item[VIRTIO_NET_F_VQ_NOTF_COAL(52)] Device supports virtqueue notification coalescing.
> > +
> >  \item[VIRTIO_NET_F_NOTF_COAL(53)] Device supports notifications coalescing.
> >  
> >  \item[VIRTIO_NET_F_GUEST_USO4 (54)] Driver can receive USOv4 packets.
> > @@ -139,6 +141,7 @@ \subsubsection{Feature bit requirements}\label{sec:Device Types / Network Device
> >  \item[VIRTIO_NET_F_NOTF_COAL] Requires VIRTIO_NET_F_CTRL_VQ.
> >  \item[VIRTIO_NET_F_RSC_EXT] Requires VIRTIO_NET_F_HOST_TSO4 or VIRTIO_NET_F_HOST_TSO6.
> >  \item[VIRTIO_NET_F_RSS] Requires VIRTIO_NET_F_CTRL_VQ.
> > +\item[VIRTIO_NET_F_VQ_NOTF_COAL] Requires VIRTIO_NET_F_CTRL_VQ.
> >  \end{description}
> >  
> >  \subsubsection{Legacy Interface: Feature bits}\label{sec:Device Types / Network Device / Feature bits / Legacy Interface: Feature bits}
> > @@ -1508,6 +1511,12 @@ \subsubsection{Control Virtqueue}\label{sec:Device Types / Network Device / Devi
> >  If the VIRTIO_NET_F_NOTF_COAL feature is negotiated, the driver can
> >  send control commands for dynamically changing the coalescing parameters.
> >  
> > +If the VIRTIO_NET_F_VQ_NOTF_COAL feature is negotiated:
> > +1. a driver can send a VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command to set the coalescing
> > +   parameters of an enabled transmit/receive virtqueue.
> > +2. a driver can send a VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET command to a device, and the device
> > +   returns the coalescing parameters of an enabled transmit/receive virtqueue.
> > +
> >  \begin{note}
> >  The behavior of the device in response to these commands is best-effort:
> >  the device may generate notifications more or less frequently than specified.
> > @@ -1519,25 +1528,55 @@ \subsubsection{Control Virtqueue}\label{sec:Device Types / Network Device / Devi
> >      le32 max_usecs;
> >  };
> >  
> > +struct virtio_net_ctrl_coal_vq {
> > +    le16 vqn;
> > +    le16 reserved;
> > +    struct virtio_net_ctrl_coal coal;
> > +};
> > +
> 
> 2 structures now, what are they doing? which command uses which?
> reader is supposed to guess and that's not good.
> 
> Also, with VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET some fields are
> output others are input. This is unusual for commands
> and needs to be documented very clearly.

Yes, this should be made clear. How do you think of the following descriptions:

"
The VIRTIO_NET_CTRL_NOTF_COAL_{TX, RX}_SET commands use the
virtio_net_ctrl_coal structure to set \field{max_usecs} and
\field{max_packets} for all transmit/receive virtqueues.

The VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command uses the
virtio_net_ctrl_coal_vq structure to set \field{max_usecs}
and \field{max_packets} for the supplied virtqueue number \field{vqn}.

The VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET command gets the values of
\field{max_usecs} and \field{max_packets} of the specified
virtqueue from the device by only setting \field{vqn} in virtio_net_ctrl_coal_vq.
"

> 
> >  #define VIRTIO_NET_CTRL_NOTF_COAL 6
> >   #define VIRTIO_NET_CTRL_NOTF_COAL_TX_SET  0
> >   #define VIRTIO_NET_CTRL_NOTF_COAL_RX_SET 1
> > + #define VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET 2
> > + #define VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET 3
> >  \end{lstlisting}
> >  
> >  Coalescing parameters:
> >  \begin{itemize}
> > +\item \field{vqn}: The virtqueue number of an enabled transmit or receive virtqueue.
> >  \item \field{max_usecs} for RX: Maximum number of microseconds to delay a RX notification.
> >  \item \field{max_usecs} for TX: Maximum number of microseconds to delay a TX notification.
> >  \item \field{max_packets} for RX: Maximum number of packets to receive before a RX notification.
> >  \item \field{max_packets} for TX: Maximum number of packets to send before a TX notification.
> >  \end{itemize}
> >  
> > -The class VIRTIO_NET_CTRL_NOTF_COAL has 2 commands:
> > +\field{vqn} points to an enabled transmit/receive virtqueue, and its value satisfies $ 0 \leq vqn < max_virtqueue_pairs \ast 2 $.
> > +
> > +\field{reserved} is reserved and it is ignored by the device.
> > +
> > +The class VIRTIO_NET_CTRL_NOTF_COAL has 4 commands:
> >  \begin{enumerate}
> > -\item VIRTIO_NET_CTRL_NOTF_COAL_TX_SET: set the \field{max_usecs} and \field{max_packets} parameters for all transmit virtqueues.
> > -\item VIRTIO_NET_CTRL_NOTF_COAL_RX_SET: set the \field{max_usecs} and \field{max_packets} parameters for all receive virtqueues.
> > +\item VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET: set the \field{max_usecs} and \field{max_packets} parameters for an enabled transmit/receive
> > +                                        virtqueue whose number is \field{vqn}.
> > +\item VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET: the device returns the \field{max_usecs} and \field{max_packets} parameters for an enabled
> > +                                        transmit/receive virtqueue whose number is \field{vqn}.
> > +\item VIRTIO_NET_CTRL_NOTF_COAL_TX_SET: have the same effect as the VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command repeated for each virtqueue of transmitq1\ldots transmitqN.
> > +\item VIRTIO_NET_CTRL_NOTF_COAL_RX_SET: have the same effect as the VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command repeated for each virtqueue of receiveq1\ldots receiveqN.
> >  \end{enumerate}
> >  
> > +If coalescing parameters are being set, the device applies the last coalescing parameters received for a
> > +virtqueue, regardless of the command used to set the parameters. For example with 2 pairs of virtqueues:
> > +# Command sequece
> > +Each of the following commands sets \field{max_usecs} and \field{max_packets} parameters for virtqueues.
> > +\begin{itemize}
> > +\item Command1: VIRTIO_NET_CTRL_NOTF_COAL_RX_SET sets parameters for virtqueue0 and virtqueue2, and, virtqueue1 and virtqueue3 retain their previous parameter values.
> > +\item Command2: VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET with \field{vqn} = 0 sets parameters for virtqueue0, and virtqueue2 retains the values from command1.
> > +\item Command3: VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET with \field{vqn}= 0, the device returns paramters of virtqueue0 set by command2.
> > +\item Command4: VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET with \field{vqn} = 1 sets parameters for virtqueue1, and virtqueue3 retains its previous values.
> > +\item Command5: VIRTIO_NET_CTRL_NOTF_COAL_TX_SET sets parameters for virtqueue1 and virtqueue3, and overrides the values set by command4.
> > +\item Command6: VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET with \field{vqn}= 1, the device returns paramters of virtqueue1 set by command5.
> > +\end{itemize}
> > +
> >  \subparagraph{Operation}\label{sec:Device Types / Network Device / Device Operation / Control Virtqueue / Notifications Coalescing / Operation}
> >  
> >  The device sends a used buffer notification once the notification conditions are met and if the notifications are not suppressed as explained in \ref{sec:Basic Facilities of a Virtio Device / Virtqueues / Used Buffer Notification Suppression}.
> > @@ -1585,14 +1624,31 @@ \subsubsection{Control Virtqueue}\label{sec:Device Types / Network Device / Devi
> >  
> >  \drivernormative{\subparagraph}{Notifications Coalescing}{Device Types / Network Device / Device Operation / Control Virtqueue / Notifications Coalescing}
> >  
> > -If the VIRTIO_NET_F_NOTF_COAL feature has not been negotiated, the driver MUST NOT issue VIRTIO_NET_CTRL_NOTF_COAL commands.
> > +If neither the VIRTIO_NET_F_NOTF_COAL nor the VIRTIO_NET_F_VQ_NOTF_COAL feature
> > +has not been negotiated, the driver MUST NOT issue VIRTIO_NET_CTRL_NOTF_COAL commands.
> 
> Triple negative? I think you mean "has been negotiated".

Yes. Thanks for pointing it out.

> 
> Generally we should consider rewriting these in a positive way
> but that's a subject for another patch.

Ok, I see.

> 
> > +
> > +A driver MUST ignore the values of coalescing paramters received from VIRTIO_NET_CTRL_NOTF_COAL_VQ_GET,
> > +if the device returns with VIRTIO_NET_ERR.
> 
> device does not return. It responds to the command.

Ok.

> 
> >  
> >  \devicenormative{\subparagraph}{Notifications Coalescing}{Device Types / Network Device / Device Operation / Control Virtqueue / Notifications Coalescing}
> >  
> >  A device SHOULD respond to the VIRTIO_NET_CTRL_NOTF_COAL commands with VIRTIO_NET_ERR if it was not able to change the parameters.
> > +Note: VIRTIO_NET_CTRL_NOTF_COAL_VQ_{SET, GET} are exception to it, which must return VIRTIO_NET_ERR if the device was unable to change the parameters.
> 
> I don't get it. what is exception to what and what's the difference,
> what does "which" refer to what return VIRTIO_NET_ERR where.
> 

Sorry for not describing clearly, please see the explanation below:
When a device receives VIRTIO_NET_CTRL_NOTF_COAL_VQ_{SET, GET} and VIRTIO_NET_CTRL_NOTF_COAL_{TX, RX}_SET respectively,
and the device was not able to change the parameters, then it behaves differently for VIRTIO_NET_CTRL_NOTF_COAL_VQ_{SET, GET} and VIRTIO_NET_CTRL_NOTF_COAL_{TX, RX}_SET, ie SHOULD and MUST.
For example:
1. VIRTIO_NET_CTRL_NOTF_COAL_RX_SET sets parameters for all transmit/receive virtqueues, but if only receiveq1 fails to be set, and receiveq2\ldots receiveqN are set successfully,
   then the device *SHOULD* respond VIRTIO_NET_ERR for VIRTIO_NET_CTRL_NOTF_COAL_RX_SET.
2. VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET sets parameters for the virtqueue whose number is \field{vqn}, if the device fails to set, the device *MUST* respond VIRTIO_NET_ERR for VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET.

> >  
> >  A device SHOULD NOT send used buffer notifications to the driver if the notifications are suppressed, even if the notification conditions are met.
> >  
> > +If a virtqueue is disabled, the device MUST clear its parameters configuration, and:
> 
> parameters here meanong coalescing parameters? It won't be clear to

Yes, I'll write clearly.

> readers who see the full spec not the patch.
> 
> > +1. if the device has received a global parameter values sent by VIRTIO_NET_CTRL_NOTF_{RX, TX}_SET:
> 
> "a" does not go with the plural "values".

Ok.

> 
> > +after the virtqueue is enabled, its parameters SHOULD be set to the global values.
> 
> 
> this is the first you mention global.

I mentioned "global" in the first sentence, but that might not seem very clear.

> 
> > +2. if the device hasn't received the VIRTIO_NET_CTRL_NOTF_{RX, TX}_SET commands:
> > +regardless of whether the device has received a VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command, after the
> > +virtqueue is enabled, its parameters SHOULD be set to 0.
> 
> Sentence too long at this point I can't say what happens after what.
> 
> 
> I feel this enumeration of cases is confusing. Would be better to find
> the logic behind all this and document.
> 
> For example, we could explain that
> 
> 
> VIRTIO_NET_CTRL_NOTF_{RX, TX}_SET set
> global values for coalescing parameters for RX and TX respectively and
> apply these global values to all VQs of the given type.
> Global values after device reset are 0.
> VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET sets the coalescing parameters for
> a given enabled VQ without changing the global values.
> Fails on a disabled VQ.  Disabling and
> re-enabling a VQ reverts its coalescing parameters back to global values.


Thanks for your example, I tried to regularize it, do you think it works?

"
The VIRTIO_NET_CTRL_NOTF_{RX, TX}_SET commands set coalescing parameters for all transmit/receive
virtqueues respectively and values of these coalescing parameters are recorded as *global values* by the device.
The device MUST set the global values of coalescing parameters to 0 after being reset.
The VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET command sets the coalescing parameters for a given enabled virtqueue without changing the global values.
After disabling and re-enabling a virtqueue, the deivce MUST revert coalescing parameters of the virtqueue back to global values.
"

> 
> 
> 
> > +
> > +When a device receives a virtqueue command to set or return coalescing parameters for the supplied \field{vqn},
> > +if the virtqueue is not enabled, the device MUST respond to the command with VIRTIO_NET_ERR.
> > +
> > +A device MAY set the coalescing parameter to a value close to a power of 2 value.
> > +
> >  Upon reset, a device MUST initialize all coalescing parameters to 0.
> >  
> >  \subsubsection{Legacy Interface: Framing Requirements}\label{sec:Device
> > -- 
> > 2.19.1.6.gb485710b