Re: [PATCH v7] virtio_net: support split header

["Michael S. Tsirkin" <mst@redhat.com>]

On Fri, Sep 09, 2022 at 03:41:54PM +0800, Heng Qi wrote:
> 
> 
> å 2022/9/5 äå4:27, Michael S. Tsirkin åé:
> > On Fri, Sep 02, 2022 at 03:36:25PM +0800, Heng Qi wrote:
> > > We need to clarify that the purpose of header splitting is to make all payloads
> > > can be independently in a page, which is beneficial for the zerocopy
> > > implemented by the upper layer.
> > absolutely, pls add motivation.
> > 
> > > If the driver does not enforce that the buffers submitted to the receiveq MUST
> > > be composed of at least two descriptors, then header splitting will become meaningless,
> > > or the VIRTIO_NET_F_SPLIT_TRANSPORT_HEADER feature should not be negotiated at this time.
> > > 
> > > 
> > > Thanks.
> > > 
> > > 
> > 
> > This seems very narrow and unecessarily wasteful of descriptors.
> > What is wrong in this:
> > 
> > <header>...<padding>... <beginning of page><data>
> > 
> > seems to achieve the goal of data in a separate page without
> > using extra descriptors.
> > 
> > thus my proposal to replace the requirement of a separate
> > descriptor with an offset of data from beginning of
> > buffer that driver sets.
> > 
> > 
> We have carefully considered your suggestion.
> 
> We refer to spec v7 and earlier as scheme A for short. Review scheme A
> below:
> 
> | receive buffer |
> 
> | 0th descriptor | 1th descriptor |
> 
> | virtnet hdr | mac | ip hdr | tcp hdr|<-- hold -->| payload |
> 
> We use a buffer plus a separate page when allocating the receive
> 
> buffer. In this way, we can ensure that all payloads can be
> 
> independently in a page, which is very beneficial for the zerocopy
> 
> implemented by the upper layer.
> 
> scheme A better solves the problem of headroom, tailroom and memory waste,
> but as you said, this solution relies on descriptor chain.
> 
> Our rethinking approach is no longer based on or using descriptor chain.
> 
> We refer to your proposed offset-based scheme as scheme B:
> 
> As you suggested, scheme B gives the device a buffer, using offset to
> indicate where to place the payload like this:
> 
> <header>...<padding>... <beginning of page><data>
> 
> But how to apply for this buffer? Since we want the payload to be placed on
> a separate page, the method we consider is to directly apply to the driver
> for two pages of contiguous memory.
> 
> Then the beginning of this contiguous memory is used to store the headroom,
> and the contiguous memory after the headroom is directly handed over to the
> device. similar to the following:
> 
> <------------------------------------------ receive buffer(2 pages)
> ----------------------------------------->
> 
> <<---------------------------------- first page
> -----------------------------------><---- second page ------>>
> 
> <<Driver reserved, the later part is filled><vheader><transport
> header>..<padding>..<beginning of page><data>>
> 
> Based on your previous suggestion, we also considered another new scheme C.
> 
> This scheme is implemented based on mergeable buffer, filling a separate
> page each time.
> 
> If the split header is negotiated and the packet can be successfully split
> by the device, the device needs to find at least two buffers, namely two
> pages, one for the virtio-net header and transport header, and the other for
> the data payload. Like the following:
> 
> | receive buffer1(page) | receive buffer2 (page) |
> 
> | virtnet hdr | mac | ip hdr | tcp hdr|<-- hold -->| payload |
> 
> At the same time, if XDP is considered, then the device needs to add
> headroom at the beginning of receive buffer1 when receiving packets, so that
> the driver can process programs similar to XDP. In order to solve this
> problem, can scheme C introduce an offset, which requires the device to
> write data from the offset position to receive buffer1, like the following:
> 
> | receive buffer (page) | receive buffer (page) |
> 
> | <-- offset(hold) --> | virtnet hdr | mac | ip hdr | tcp hdr|<-- hold -->|
> payload |

And in fact, B and C both use an offset now, right?

> Then we simply compare the advantages and disadvantages of scheme A(spec
> v7), scheme B (offset buffer(2 pages)) and scheme C (based on mergeable
> buffer):
> 
> 1. desc chain:
> 
> - A depends on desciptor chain; - B, C do not depend on desciptor chain.
> 
> 2. page alloc
> 
> - B fills two consecutive pages, which causes a great waste of memory for
> small packages such as arp; - C fills a single page, slightly better than B.
> 
> 3. Memory waste:
> 
> - The memory waste of scheme A is mainly the 0th descriptor that is skipped
> by the device;

there's also the cost of the indirect buffer since that is used when
there is a chain.

> - When scheme B and scheme C successfully split the header,
> there is a huge waste of the first page, but the first page can be quickly
> released by copying.
> 
> 4. headroom
> 
> - The headrooms of plan A and plan B are reserved; - Scheme C requires the
> driver to set off to let the device skip off when using receive buffer1.
> 
> 5. tailroom
> 
> - When splitting the header, skb usually needs to store each independent
> page in the non-linear data area based on shinfo. - The tailroom of scheme A
> is reserved by itself; - Scheme B requires the driver to set the reserved
> padding area for the first receive buffer(2 pages) to use shinfo when the
> split header is not successfully executed; - Scheme C requires the driver to
> set max_len for the first receive buffer(page).
> 
> 
> Which plan do you prefer?

I think either both B and C depending on the mergeable buffers flag,
or just one of these two.

> ---
> 
> Thanks.