Date:   Mon, 10 Feb 2020 17:21:11 +0800
From:   "Liu, Jing2" <jing2.liu@...ux.intel.com>
To:     Stefan Hajnoczi <stefanha@...hat.com>
Cc:     virtio-dev@...ts.oasis-open.org, slp@...hat.com,
        linux-kernel@...r.kernel.org,
        Chao Peng <chao.p.peng@...ux.intel.com>,
        Zha Bin <zhabin@...ux.alibaba.com>,
        Liu Jiang <gerry@...ux.alibaba.com>
Subject: Re: [virtio-dev][PATCH v1 1/2] virtio-mmio: Add MSI and different
 notification address support


On 1/23/2020 12:56 AM, Stefan Hajnoczi wrote:
> On Mon, Jan 13, 2020 at 07:54:06PM +0800, Liu, Jing2 wrote:
>>>>>>         \end{note}
>>>>>>       }
>>>>>>       \hline
>>>>>> @@ -1671,25 +1671,23 @@ \subsection{MMIO Device Register Layout}\label{sec:Virtio Transport Options / Vi
>>>>>>         accesses apply to the queue selected by writing to \field{QueueSel}.
>>>>>>       }
>>>>>>       \hline
>>>>>> -  \mmioreg{QueueNotify}{Queue notifier}{0x050}{W}{%
>>>>>> -    Writing a value to this register notifies the device that
>>>>>> -    there are new buffers to process in a queue.
>>>>>> +  \mmioreg{QueueNotify}{Queue notifier}{0x050}{RW}{%
>>>>>> +    Reading from the register returns the virtqueue notification configuration.
>>>>>> -    When VIRTIO_F_NOTIFICATION_DATA has not been negotiated,
>>>>>> -    the value written is the queue index.
>>>>>> +    See \ref{sec:Virtio Transport Options / Virtio Over MMIO / MMIO-specific Initialization And Device Operation / Notification Address}
>>>>>> +    for the configuration format.
>>>>>> -    When VIRTIO_F_NOTIFICATION_DATA has been negotiated,
>>>>>> -    the \field{Notification data} value has the following format:
>>>>>> +    Writing when the notification address calculated by the notification configuration
>>>>>> +    is just located at this register.
>>>>> I don't understand this sentence.  What happens when the driver writes
>>>>> to this register?
>>>> We're trying to define the notification mechanism that, driver MUST read
>>>> 0x50 to get the notification configuration
>>>>
>>>> and calculate the notify address. The writing case here is that, the notify
>>>> address is just located here e.g. notify_base=0x50, notify_mul=0.
>>> I still don't understand what this means.  It's just an English issue
>>> and it will become clear if you can rephrase what you're saying.
>> Sure, let me try to explain it. :)
>>
>> The different notification locations are calculated via the structure
>> returned by reading this register.
>>
>> le32 {
>>      notify_base : 16;
>>      notify_multiplier : 16;
>> };
>>
>> location=notify_base + queue_index * notify_multiplier
>>
>> The location might be the same when mul=0, and furthermore, it might be
>> equal to 0x50 (notify_base=0x50, notify_mul=0) so we make this register W
>> too.
>>
>> So we said, the register is RW and W is only for such scenario.
>>
>> Feel free to tell me if it's still confusing.
> I understand now:
>
>    Devices that only require a single notify address may set
>    notify_base=0x50 and notify_multiplier=0 to use the Queue Notifier
>    register itself for notifications.  In this case the driver writes to
>    Queue Notifier to notify the device that there are new buffers in a
>    virtqueue.
>
> Perhaps you could include this in the text.

Thanks for the guide. Since v2 was sent out, we'll add such text in 
later version.

Jing

>>>>>> -    \lstinputlisting{notifications-le.c}
>>>>>> -
>>>>>> -    See \ref{sec:Virtqueues / Driver notifications}~\nameref{sec:Virtqueues / Driver notifications}
>>>>>> -    for the definition of the components.
>>>>>> +    See \ref{sec:Virtio Transport Options / Virtio Over MMIO / MMIO-specific Initialization And Device Operation / Available Buffer Notifications}
>>>>>> +    to see the notification data format.
>>>>>>       }
>>>>>>       \hline
>>>>>>       \mmioreg{InterruptStatus}{Interrupt status}{0x60}{R}{%
>>>>>>         Reading from this register returns a bit mask of events that
>>>>>> -    caused the device interrupt to be asserted.
>>>>>> +    caused the device interrupt to be asserted. This is only used
>>>>>> +    when MSI is not enabled.
>>>>>>         The following events are possible:
>>>>>>         \begin{description}
>>>>>>           \item[Used Buffer Notification] - bit 0 - the interrupt was asserted
>>>>>> @@ -1703,7 +1701,7 @@ \subsection{MMIO Device Register Layout}\label{sec:Virtio Transport Options / Vi
>>>>>>       \mmioreg{InterruptACK}{Interrupt acknowledge}{0x064}{W}{%
>>>>>>         Writing a value with bits set as defined in \field{InterruptStatus}
>>>>>>         to this register notifies the device that events causing
>>>>>> -    the interrupt have been handled.
>>>>>> +    the interrupt have been handled. This is only used when MSI is not enabled.
>>>>>>       }
>>>>>>       \hline
>>>>>>       \mmioreg{Status}{Device status}{0x070}{RW}{%
>>>>>> @@ -1762,6 +1760,31 @@ \subsection{MMIO Device Register Layout}\label{sec:Virtio Transport Options / Vi
>>>>>>         \field{SHMSel} is unused) results in a base address of
>>>>>>         0xffffffffffffffff.
>>>>>>       }
>>>>>> +  \hline
>>>>>> +  \mmioreg{MsiStatus}{MSI status}{0x0c0}{R}{%
>>>>>> +    Reading from this register returns the global MSI enable/disable status and maximum
>>>>>> +    number of virtqueues that device supports.
>>>>>> +    \lstinputlisting{msi-status.c}
>>>>>> +  }
>>>>> Why is it necessary to combine the number of virtqueues and global
>>>>> MSI enable/disable into a single 16-bit field?
>>>> Originally, we want this 16-bit Read-Only, so we put some RO things together
>>>> and separate
>>>>
>>>> enable setting command to next register.
>>>>
>>>>> virtio-mmio uses 32-bit registers.  It doesn't try hard to save register
>>>>> space so it's strange to do it here (11-bit number of virtqueue field
>>>>> but 32-bit QueueSel field).
>>>> In order to improve performance/save register space,  we combine some data
>>>> together.
>>>>
>>>> For example, combine MSI cmd operator (e.g. enable/disable, vector setup)
>>>> with argument (e.g. 1/0,  queue index).
>>>>
>>>> But it seems we miss the consistency with QueueSel.  So do you think if the
>>>> max queue number should be 32-bit,
>>>>
>>>> which means it must be the same with QueueSel? If so, I guess we need some
>>>> re-organization. :)
>>> I suggest following the 32-bit register size convention unless there is
>>> a specific reason why using other register sizes is absolutely necessary.
>> Yes, let's keep consistency with QueueSel and re-organize other registers.
>>
>> I feel concern why Available Buffer Notifcations (section describing
>> VIRTIO_F_NOTIFICATION_DATA) makes vq index as 16bit?
> As you mentioned, the valid range of virtqueue numbers is only 16 bits
> due to non-MMIO parts of the specification using 16 bits.
>
> However, I think it makes sense to stick to the MMIO transport 32-bit
> register size convention for consistency.  Devices just won't support
> values above 0xffff.
>
> Stefan

Powered by blists - more mailing lists