qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [virtio-dev][RFC v3] virtio-sdm: new device specificati


From: Christian Pinto
Subject: Re: [Qemu-devel] [virtio-dev][RFC v3] virtio-sdm: new device specification
Date: Mon, 12 Sep 2016 17:42:18 +0200
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Thunderbird/45.2.0

Hello Edgar,


sorry for the delay.


On 08/09/2016 17:38, Edgar E. Iglesias wrote:
On Thu, Sep 08, 2016 at 09:06:00AM +0200, Christian Pinto wrote:
Hello Edgar,


On 07/09/2016 18:02, Edgar E. Iglesias wrote:
On Wed, Sep 07, 2016 at 05:39:08PM +0200, Christian Pinto wrote:
On 07/09/2016 09:51, Edgar E. Iglesias wrote:
On Wed, Sep 07, 2016 at 09:24:39AM +0200, Christian Pinto wrote:
Hello Edgar,

thanks for your comments.

Thanks for the clarification, I have a few follow-up questions/comments.
On 06/09/2016 23:43, Edgar E. Iglesias wrote:
Hi,

Sorry for the delay. I have a few questions.

I don't fully understand the purpose of this. Could you elaborate a little on 
that?
You need a real IPI signal to drive this I guess, so it's a little bit of a
chicken and egg problem.
Usually in AMP-like systems CPUs signal each other with IPIs by exploiting
a dedicated hardware peripheral, like a hardware mailbox. So I see the
concept of the SDM (master, slave and communication channels) as the
hardware device to actually implement and deliver the IPIs.
In this case the SDM is designed in a way to be able to deliver multiple
signals (e.g., BOOT, RESET) and not just a single interrupt.

A useful feature I can see is multiplexing a single underlying IPI (driving
an sdm channel) into multiple "virtual" IPI signals. Is that the main
use-case?
In the SDM multiple slaves can connect to the same SDM channel (e.g.,
socket)
and the master uses the channel to deliver the signal (e.g., IRQ) to the
destination slave device. The other way around is also possible, slaves
signaling
the master.
I guess this is the multiplexing features you were talking about.
Right, but for this to be more useful, I think you should pass on some
payload with the IRQ signal. That will allow this channel to be used
to enable notification mechanisms for many other channels over a single
HW irq (e.g the payload would be like the irq vector or irq line
All signals carry a 64 bits payload that for the time being is ignored
for the IRQ signal. I now see your point, thanks for the clarification.
The payload could definitely be used as an IRQ vector, so that SW
can register to a specific IRQ line of the SDM HW interrupt.
Thanks Christian,

Perhaps we could update the spec and elaborate a little on this and
change the payload from ignored to provide and notifier/irq index?

Sure. I will improve this part in the next release.

Another question:
For the reverese path, I don't see how it will be reasonably implemented
in the remote cpus various sw stacks sharing a single virtio queue.
It would seem like you would need a reverese channel per remoteproc
to avoid locking (very messy in AMP systems) to safely update
the reverese gh_vq queue?

Or how do you see multiple AMP systems safely updating a single virtio
queue in parallel?
The AMP systems are not all interacting with the same virtio queues.
As described in the specification the device can be a master or a slave.
The master processor will interact with a master SDM instance,
while remote processors with slave SDM instances.
Master/slave instances should be seen as multiple ports of the same
SDM.
Master and slave SDM instances communicate using the SDM Channel.
So, no need for locking or to handle concurrent updates from
multiple processors to the same virtio queue.

These are the parts that I find a little confusing:

\begin{description}
\item[0] hg_vq
\item[1] gh_vq
\end{description}

Queue 0 is used for device-to-driver communication (i.e., notification of a
signal), while queue 1 for driver-to-device communication.

...

The master slave behavior is meant on purpose to reflect the AMP like type of 
communication
where a master processor controls one or more slave processors.
A master SDM instance can send a signal to any of the slaves instances,
while slaves can only signal the master.

...

The \texttt{SDMSignalData} structure is first filled by the source SDM kernel
virtio driver and sent over the gh_vq.


I interpret that as there being a single gh_vq queue and multiple slaves 
sending events
on that queue. Do you mean that there are multiple gh_vq's?
Am I missing something?
When I talk about master/slave instance I mean two different SDM
devices, one with master and the other with slave behavior.
This means that each SDM device will have its own pair of virtio
queues (gh_vq and hg_vq). Each remote processor (or slave)
will use its own SDM slave device and update only the related virtio
queues. The same is true for the master.
All the instances of the SDM interact between each other through
the SDM Channel (e.g., shared memory or network sockets in case
multiple nodes are involved in the virtualization of the AMP).

So, as an example, if your system has 1 master processor and 2
remote processors, you will have to instantiate three SDM devices
(1 master and 2 slaves). This means three different virtio devices,
each one interacting with a different processor.

Probably this aspect is not explicitly highlighted in the text, and I
can clarify it a bit in the next release.
Thanks Christian,

Yes, I think this part is confusing in that aspect:

"The master slave behavior is meant on purpose to reflect the AMP like type of 
communication
where a master processor controls one or more slave processors.
A master SDM instance can send a signal to any of the slaves instances,
while slaves can only signal the master."

OK probably here I can explicitly describe that the signal is sent over the
communication channel, and that each device instance (either master or slave)
has its own pair of virtio queues shared with the driver.

As an example the flow a signal from the master to a slave is the following:

master driver -> vqueues1 -> device1 -> comm channel -> device2 -> vqueues2 -> slave driver

Where device1 is the master instance and device2 the slave.
On the right the fan-out of the communication channel is the number of
slaves (one in this example).


I agree that it makes more sense to have individual (point to point) sdm 
devices.

Another question, how does the following stuff fit in if we now agree that we
want point-to-point (single master to single slave) SDM instances?

Actually it is a 1-to-N type of communication. Each master can be connected to multiple slave SDM devices, and it is the job of the communication channel to route requests to the correct destination slave. The opposite is true to send
a signal from a slave to the master.
So the data structure below is needed by the master to identify the actual
number of slaves vs the maximum expected.

Christian
\begin{lstlisting}
struct virtio_sdm_config {
        u16   max_slaves;
        u16   current_slaves;
        u32   device_id;
};
\end{lstlisting}

Cheers,
Edgar


Please let me know whether you think that further information is
needed to make this part of the spec 100% clear.

Thanks,

Christian
Best regards,
Edgar


Christian

Best regards,
Edgar


However it is worth to notice that one SDM channel is not dedicated to one
specific signal, but rather it is used to deliver all the signals
implemented
by the device.

Regarding the boot and reset. Typically there's a reset signal you release and
the remote CPU starts running. You can ofcourse reset the CPU and restart
at anytime. Not sure if you need an additional BOOT virtual signal.

There may also be additional mechanisms to control the startup address of the 
remote CPU.
Any thoughts on that?
The purpose of the BOOT signal is also to pass the remote CPU its startup
address. So the idea is to use the payload of the signal to send the slave
processor
its startup address over the SDM channel and SDM slave device. Our
implementation
for QEMU uses the BOOT signal for this purpose.

This part of the semantic of the BOOT signal is not described in the
documentation
since I believe it is a choice of the actual implementation of the SDM
whether to
pass the startup address or not (i.e. there might be a platform where the
boot
address of slave processors is somehow hard-coded).
I can add the explanation to the device specification to make it more clear.


Thanks,

Christian

Best regards,
Edgar


On Tue, Aug 30, 2016 at 01:22:26PM +0200, Christian Pinto wrote:
Hello,

are there any comments?


Christian


On 09/08/2016 09:37, Christian Pinto wrote:
This patch adds the specification of the Signal Dristribution Module virtio
device to the current virtio specification document.

Signed-off-by: Christian Pinto<address@hidden>
<mailto:address@hidden>
Signed-off-by: Baptiste Reynal<address@hidden>
<mailto:address@hidden>

---
v2 -> v3:
- Removed master field from configuration and replaced with device_id
- Added new RESET signal
- Added signals definition into device specs
- Added feature bits associated to signals

v1 -> v2:
- Fixed some typos
- Removed dependencies from QEMU
- Added explanation on how SDM can be used in AMP systems
- Explained semantics of payload field in SDMSignalData struct
---
  content.tex    |   2 +
  virtio-sdm.tex | 166 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
  2 files changed, 168 insertions(+)
  create mode 100644 virtio-sdm.tex

diff --git a/content.tex b/content.tex
index 3317916..7fcf779 100644
--- a/content.tex
+++ b/content.tex
@@ -5643,6 +5643,8 @@ descriptor for the \field{sense_len}, \field{residual},
  \field{status_qualifier}, \field{status}, \field{response} and
  \field{sense} fields.
+\input{virtio-sdm.tex}
+
  \chapter{Reserved Feature Bits}\label{sec:Reserved Feature Bits}
  Currently there are three device-independent feature bits defined:
diff --git a/virtio-sdm.tex b/virtio-sdm.tex
new file mode 100644
index 0000000..ee43e23
--- /dev/null
+++ b/virtio-sdm.tex
@@ -0,0 +1,166 @@
+\section{Signal Distribution Module}\label{sec:Device Types / SDM Device}
+
+The virtio Signal Distribution Module is meant to enable Inter Processor signal
+exchange.
+An example are Inter Processor Interrupts used in AMP systems, for the
+processors involved to notify the presence of new data in the communication
+queues.
+In AMP systems signals are used for various purposes, an example are remoteproc
+or RPMSG. In the former signals are used by a master processor to trigger
+the boot of a slave processor. While the latter, RPMSG, uses virtio queues as a
+message exchange medium between processors. In this case the SDM can be used to
+generate the interrupt associated to the kick of a virtio queue.
+
+There are three signal types supported by the device, namely the
+\textit{IRQ} signal, \textit{BOOT} signal and \textit{RESET} signal. Such set 
of
+signals covers most of the needs of an AMP system, where a master processor can
+trigger the boot or reset of slave processors, and processors send IRQs between
+each other.
+
+\subsection{Device ID}\label{sec:Device Types / SDM Device / Device ID}
+
+21
+
+\subsection{Virtqueues}\label{sec:Device Types / SDM Device / Virtqueues}
+
+\begin{description}
+\item[0] hg_vq
+\item[1] gh_vq
+\end{description}
+
+Queue 0 is used for device-to-driver communication (i.e., notification of a
+signal), while queue 1 for driver-to-device communication.
+
+\subsection{Feature bits}\label{sec:Device Types / SDM Device / Feature bits}
+
+\begin{description}
+\item[VIRTIO_SDM_F_IRQ_SIG (0)] Device supports the IRQ signal.
+
+\item[VIRTIO_SDM_F_BOOT_SIG (1)] Device supports the BOOT signal.
+
+\item[VIRTIO_SDM_F_RESET_SIG (2)] Device supports the RESET signal.
+\end{description}
+
+This set of bits is used by each virtio SDM device to declare which of the
+signals it supports. By specification each device can support all or a subset 
of
+the defined signals.
+
+\subsection{Device configuration layout}\label{sec:Device Types / SDM Device /
+Device configuration layout}
+
+The device configuration is composed by three fields:
+\texttt{max_slaves}, \texttt{current_slaves} and the \texttt{device_id}.
+
+\begin{lstlisting}
+struct virtio_sdm_config {
+       u16   max_slaves;
+       u16   current_slaves;
+       u32   device_id;
+};
+\end{lstlisting}
+
+The SDM device can be instantiated either as a master or as a slave. The master
+slave behavior is meant on purpose to reflect the AMP like type of 
communication
+where a master processor controls one or more slave processors.
+A master SDM instance can send a signal to any of the slaves instances,
+while slaves can only signal the master.
+
+The \texttt{master} field of the configuration space is meant to be read by the
+driver to figure out whether a specific instance is a master or a slave.
+The \texttt{max_slaves} field contains the maximum number of slaves supported 
by
+the SDM device. A configuration change notification is sent to the driver each
+time the value of \texttt{max_slaves} is changed from the device side.
+Finally, the \texttt{current_slaves} field contains the actual number of slaves
+registered to the master SDM. This field is a read only field. Finally the
+\texttt{device_id} field is used by each driver to know the ID of the device it
+is attached to, the field contains 0 in case of a master instance. A driver
+figures out whether it is attached to a master or a slave instance thanks to 
this
+field.
+
+\subsection{Device Initialization}\label{sec:Device Types / SDM Device /
+evice Initialization}
+
+During initialization the \texttt{hg_vq} and \texttt{gh_vq} are identified and
+the device is immediately operational. A master driver instance can access the
+number of slaves registered at any time by reading the configuration space of
+the device.
+
+During the initialization phase the device connects also to the communication
+channel. It has to be noted that the behavior of the device is
+independent from the communication channel used, that is a detail of each
+specific implementation of the SDM device.
+
+The last phase of the initialization is the negotiation of the feature bits.
+Each device implementation declares the signals supported by offering all or a
+subset of the three feature bits (VIRTIO_SDM_F_IRQ_SIG, VIRTIO_SDM_F_BOOT_SIG,
+VIRTIO_SDM_F_RESET_SIG). The SDM driver will be aware of the set of signals to
+handle thanks to this negotiation phase.
+
+\subsection{Device Operation}\label{sec:Device Types / SDM Device / Device
+peration}
+
+The SDM device handles signals coming from the two following sources:
+
+\begin{enumerate}
+\item The local processor to which the device is attached to.
+\item The communication channel connecting to other slaves/master.
+\end{enumerate}
+
+The first case is verified when the processor attached to the SDM device wants
+to send a signal to a second SDM device instance.
+The second case is instead when an SDM device instance receives a signal from
+another SDM device, to be forwarded to the local processor.
+It is important to note that due to the master/slave behavior, slaves cannot
+signal among themselves but only with the master SDM instance.
+
+In both cases, before sending over the communication channel, the signal is
+packed in the \texttt{SDMSignalData} data structure.
+
+\begin{lstlisting}
+enum sdm_signal_type {
+    SDM_IRQ,
+    SDM_BOOT,
+    SDM_RESET,
+};
+
+struct SDMSignalData {
+    uint32_t type;
+    uint32_t slave;
+    uint32_t payload[2];
+};
+\end{lstlisting}
+
+The \texttt{type} field indicates the type of signal to be sent to the
+destination SDM. The current implementation supports three signal types.
+The \texttt{SDM_IRQ} signal is used to send an inter processor interrupt, while
+the \texttt{SDM_BOOT} signal is sent to trigger the boot of the destination
+processor. The boot signal is meant to be used in an AMP like scenario where a
+master processor boots one or more slave processors (e.g., via remoteproc).
+The \texttt{SDM_RESET} is also meant to be used in AMP like scenarios, to
+trigger the reset of the target slave processor. As an assumption a driver
+cannot trigger the reset of the device it is attached to, so each driver
+implementation should ignore reset signals where the source slave corresponds 
to
+the device ID the driver is attached to.
+This is done by comparing, when a message is recevied, the value of
+the \texttt{slave} field of the \texttt{SDMSignalData} data structure with the
+\texttt{device_id} field of the configuration space.
+The \texttt{slave} field contains the id of the SDM instance destination of the
+signal. Id 0 is reserved for the master, from 1 onwards for the slaves.
+This means that the \texttt{slave} field will always contain 0 when the source
+of the signal is a slave SDM instance, while the actual id of the slave in case
+of a master. When instead a message is recevied, this field contains the ID of
+the slave source of the signal.
+The \texttt{payload} is used to pass extra accompanying information with the
+signal.
+The semantics of the payload field depends on the signal itself. In case of
+\texttt{SDM_IRQ} signal, the payload field is ignored since interrupts do not
+need any extra information to be handled. In the case of \texttt{SDM_BOOT}
+signal the payload contains the boot address of the slave processor, to be used
+at the destination to initialize the program counter register before the actual
+boot process is started. Finally the payload field is ignored also in case of
+\texttt{SDM_RESET} signal, since no extra information is needed to trigger the
+reset of the target processor.
+
+
+The \texttt{SDMSignalData} structure is first filled by the source SDM kernel
+virtio driver and sent over the gh_vq.




reply via email to

[Prev in Thread] Current Thread [Next in Thread]