RE: [virtio-comment] Re: [PATCH v2] virtio-net: Add an optional device control over the receive buffers length

[Vitaly Mireyno <vmireyno@marvell.com>]

>> >
>> >I still think this is too restrictive. Switching between e.g.
>> >XDP/non-XDP workloads demands ability to change this without device reset.
>> >I am thinking about sending the avail index to device with a command
>> >to change the size, this way device can figure out where are the old
>> >and where are the new buffers.
>> >Doesn't look too hard on the hardware, does it?
>> >
>>
>> Changing descriptors size once in a while sounds reasonable and possible.
>> I see a problem though with this specific method - updating the rx descriptors size is an asynchronous
>slow-path operation (in vDPA it would go through a SW emulation). So there is no guarantee that the
>ring index, the driver requests, will be delivered on time to the device.
>> I can suggest another method, where the driver requests to change a descriptor size through a slow-
>path, but fast-path synchronization will come from the device - it will set a dedicated flag in the first
>used rx descriptor with the new size (from the device perspective).
>
>Now I'm confused.  The descriptor length is *already* in the descriptor.
>Why do we need a new flag?
>

If I understand correctly, the device will put a received packet length into the descriptor's 'length' field. So the driver will have no way of knowing when the device switches to the new descriptor length. More details below.

>
>> So the flow of changing the descriptor size could be as follows:
>>
>> For enlarging descriptors size:
>>  1. Driver enlarges descriptors size. All new descriptors are made available with the new size.
>Obviously, the buffers behind the descriptors must be enlarged as well.
>>  2. Driver requests the device to change descriptors size (by writing to the device configuration
>space). At this stage the device still thinks that all descriptors are of an old (smaller) size. Although new
>descriptors are of a new size, the driver must treat them as if they were of an old size (i.e. the device
>would not fill the whole descriptor).
>>  3. At some point, the device receives the driver's request, and it changes internally the rx descriptors
>size. The device sets a new_size flag in the next descriptor used.
>>  4. When the driver receives a used descriptor with a new_size flag set, from that point on it can treat
>all descriptors as being of a new size.
>
>So the only thing that happens with the flag is it's copied from source to destination right? I guess in
>that case, just using descriptor ID is enough.
>

I'm not sure I understand what do you mean by "copied from source to destination".

When driver wants to change the descriptor length from X to Y (Y > X), there will be a period of time when descriptors' length is Y, but the device thinks it's X. For packets that are smaller than X, it doesn't really matter - the device will use a single descriptor per packet, and write out its length. For packets larger than X, the device will split the packet to buffers pointed by multiple descriptors, but each buffer will be filled with up to X bytes (and not Y).
Currently, there is a requirement in the spec, that "If a receive packet is spread over multiple buffers, the device MUST use all buffers but the last (i.e. the first num_buffers-1 buffers) completely up to the full length of each buffer supplied by the driver.".
So we can tell the driver that till it sees the new_size flag, it should treat used descriptors length as if it was X and not Y.
Alternatively, perhaps the requirement can be changed, such that driver will always read the used descriptor's length, and will not assume that it was fully populated. In this case a new flag will not be needed.

When Y < X, there will be a period of time when the device thinks that the descriptors' length is X, so the driver can't allocate smaller buffers. It must continue providing descriptors pointing to large buffers, till the device signals that it has received the request, and from now on it will start using smaller buffers. A new_size flag can be such a signal. From the moment the driver sees a new_size flag set in a used descriptor, it can start allocating small buffers, and make new available descriptors with length Y. However, there could still be outstanding available descriptors with length X, so the driver must treat them as if they were of length Y (like in the previous case).


>> For reducing descriptors size:
>>  1. Driver requests the device to change descriptors size.
>>  2. At some point, the device receives the driver's request, and it changes internally the rx descriptors
>size. the device sets a new_size flag in the next descriptor used.
>>  3. When the driver receives a used descriptor with a new_size flag set, from that point on it must
>treat all the descriptors as being of the new size.
>>  4. Now the driver can actually reduce descriptors size. All new descriptors will be made available with
>a new size.
>
>Except that we should be talking about buffers not descriptors, sounds ok. Fundamentally, the main
>rule we are relaxing is that device uses a whole buffer. Need to update the relevant spec part, making
>an exception and link to this process.
>

I was mostly talking about descriptors and not buffers, because this is what exposed to the device. As far as I understand, the device doesn't care about the buffer length, but only about the descriptor 'length' field.