[Xen-devel] [PATCH v2 0/1] netif: staging grants for I/O requests

[Joao Martins <joao.m.martins@xxxxxxxxxx>]

Hey,

This is v2 taking into consideration all comments received from RFC/v1.
The approach has significantly changed and we use control ring messages to
the manage the permanent gref mappings - as also presented at XDDS17. The
specification is right after the diffstat. I kept the same name "staging
grants" but feel free to suggest an alternative.

Reference implementation also here (on top of net-next):

https://github.com/jpemartins/linux.git xen-net-stg-gnts-v2

I also would like to note that this lays some groundwork for netback to
work with a known set of pages, hence there will be less effort involved
into making netback work with something like zerogrant.

Cheers,
Joao

Joao Martins (1):
  public/io/netif.h: add gref mapping control messages

 xen/include/public/io/netif.h | 114 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 114 insertions(+)
---
% Staging grants for network I/O requests
% Joao Martins <<joao.m.martins@xxxxxxxxxx>>
% Revision 2

\clearpage

--------------------------------------------------------------------
Architecture(s): Any
--------------------------------------------------------------------

# Background and Motivation

At the Xen hackaton '16 networking session, we spoke about having a permanently
mapped region to describe header/linear region of packet buffers. This document
outlines the proposal covering motivation of this and applicability for other
use-cases alongside the necessary changes. This proposal is an RFC and also
includes alternative solutions.

The motivation of this work is to eliminate grant ops for packet I/O intensive
workloads such as those observed with smaller requests size (i.e. <= 256 bytes
or <= MTU). Currently on Xen, only bulk transfer (e.g. 32K..64K packets) are the
only ones performing really good (up to 80 Gbit/s in few CPUs), usually
backing end-hosts and server appliances. Anything that involves higher packet
rates (<= 1500 MTU) or without sg, performs badly almost like a 1 Gbit/s
throughput.

# Proposal

The proposal is to leverage the already implicit copy from and to packet linear
data on netfront and netback, to be done instead from a permanently mapped
region. In some (physical) NICs this is known as header/data split.

Specifically some workloads (e.g. NFV) it would provide a big increase in
throughput when we switch to (zero)copying in the backend/frontend, instead of
the grant hypercalls. Thus this extension aims at futureproofing the netif
protocol by adding the possibility of guests setting up a list of grants that
are set up at device creation and revoked at device freeing - without taking
too much grant entries in account for the general case (i.e. to cover only the
header region <= 256 bytes, 16 grants per ring) while configurable by kernel
when one wants to resort to a copy-based as opposed to grant copy/map.

\clearpage

# General Operation

Here we describe how netback and netfront general operate, and where the 
proposed
solution will fit. The security mechanism currently involves grants references
which in essence are round-robin recycled 'tickets' stamped with the GPFNs,
permission attributes, and the authorized domain:

(This is an in-memory view of struct grant_entry_v1):

     0     1     2     3     4     5     6     7 octet
    +------------+-----------+------------------------+
    | flags      | domain id | frame                  |
    +------------+-----------+------------------------+

Where there are N grant entries in a grant table, for example:

    @0:
    +------------+-----------+------------------------+
    | rw         | 0         | 0xABCDEF               |
    +------------+-----------+------------------------+
    | rw         | 0         | 0xFA124                |
    +------------+-----------+------------------------+
    | ro         | 1         | 0xBEEF                 |
    +------------+-----------+------------------------+

      .....
    @N:
    +------------+-----------+------------------------+
    | rw         | 0         | 0x9923A                |
    +------------+-----------+------------------------+

Each entry consumes 8 bytes, therefore 512 entries can fit on one page.
The `gnttab_max_frames` which is a default of 32 pages. Hence 16,384
grants. The ParaVirtualized (PV) drivers will use the grant reference (index
in the grant table - 0 .. N) in their command ring.

\clearpage

## Guest Transmit

The view of the shared transmit ring is the following:

     0     1     2     3     4     5     6     7 octet
    +------------------------+------------------------+
    | req_prod               | req_event              |
    +------------------------+------------------------+
    | rsp_prod               | rsp_event              |
    +------------------------+------------------------+
    | pvt                    | pad[44]                |
    +------------------------+                        |
    | ....                                            | [64bytes]
    +------------------------+------------------------+-\
    | gref                   | offset    | flags      | |
    +------------+-----------+------------------------+ +-'struct
    | id         | size      | id        | status     | | netif_tx_sring_entry'
    +-------------------------------------------------+-/
    |/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/| .. N
    +-------------------------------------------------+
 
Each entry consumes 16 octets therefore 256 entries can fit on one page.`struct
netif_tx_sring_entry` includes both `struct netif_tx_request` (first 12 octets)
and `struct netif_tx_response` (last 4 octets).  Additionally a `struct
netif_extra_info` may overlay the request in which case the format is:

    +------------------------+------------------------+-\
    | type |flags| type specific data (gso, hash, etc)| |
    +------------+-----------+------------------------+ +-'struct
    | padding for tx         | unused                 | | netif_extra_info'
    +-------------------------------------------------+-/

In essence the transmission of a packet in a from frontend to the backend
network stack goes as following:

**Frontend**

1) Calculate how many slots are needed for transmitting the packet.
   Fail if there are aren't enough slots.

[ Calculation needs to estimate slots taking into account 4k page boundary ]

2) Make first request for the packet.
   The first request contains the whole packet size, checksum info,
   flag whether it contains extra metadata, and if following slots contain
   more data.

3) Put grant in the `gref` field of the tx slot.

4) Set extra info if packet requires special metadata (e.g. GSO size)

5) If there's still data to be granted set flag `NETTXF_more_data` in
request `flags`.

6) Grant remaining packet pages one per slot. (grant boundary is 4k)

7) Fill resultant grefs in the slots setting `NETTXF_more_data` for the N-1.

8) Fill the total packet size in the first request.

9) Set checksum info of the packet (if the chksum offload if supported)

10) Update the request producer index (`req_prod`)

11) Check whether backend needs a notification

11.1) Perform hypercall `EVTCHNOP_send` which might mean a __VMEXIT__
      depending on the guest type.

**Backend**

12) Backend gets an interrupt and runs its interrupt service routine.

13) Backend checks if there are unconsumed requests

14) Backend consume a request from the ring

15) Process extra info (e.g. if GSO info was set)

16) Counts all requests for this packet to be processed (while
`NETTXF_more_data` is set) and performs a few validation tests:

16.1) Fail transmission if total packet size is smaller than Ethernet
minimum allowed;

  Failing transmission means filling `id` of the request and
  `status` of `NETIF_RSP_ERR` of `struct netif_tx_response`;
  update rsp_prod and finally notify frontend (through `EVTCHNOP_send`).

16.2) Fail transmission if one of the slots (size + offset) crosses the page
boundary

16.3) Fail transmission if number of slots are bigger than spec defined
(18 slots max in netif.h)

17) Allocate packet metadata

[ *Linux specific*: This structure emcompasses a linear data region which
generally accomodates the protocol header and such. Netback allocates up to 128
bytes for that. ]

18) *Linux specific*: Setup up a `GNTTABOP_copy` to copy up to 128 bytes to 
this small
region (linear part of the skb) *only* from the first slot.

19) Setup GNTTABOP operations to copy/map the packet

20) Perform the `GNTTABOP_copy` (grant copy) and/or `GNTTABOP_map_grant_ref`
    hypercalls.

[ *Linux-specific*: does a copy for the linear region (<=128 bytes) and maps the
         remaining slots as frags for the rest of the data ]

21) Check if the grant operations were successful and fail transmission if
any of the resultant operation `status` were different than `GNTST_okay`.

21.1) If it's a grant copying backend, therefore produce responses for all the
the copied grants like in 16.1). Only difference is that status is
`NETIF_RSP_OKAY`.

21.2) Update the response producer index (`rsp_prod`)

22) Set up gso info requested by frontend [optional]

23) Set frontend provided checksum info

24) *Linux-specific*: Register destructor callback when packet pages are freed.

25) Call into to the network stack.

26) Update `req_event` to `request consumer index + 1` to receive a notification
    on the first produced request from frontend.
    [optional, if backend is polling the ring and never sleeps]

27) *Linux-specific*: Packet destructor callback is called.

27.1) Set up `GNTTABOP_unmap_grant_ref` ops for the designated packet pages.

27.2) Once done, perform `GNTTABOP_unmap_grant_ref` hypercall. Underlying
this hypercall a TLB flush of all backend vCPUS is done.

27.3) Produce Tx response like step 21.1) and 21.2)

[*Linux-specific*: It contains a thread that is woken for this purpose. And
it batch these unmap operations. The callback just queues another unmap.]

27.4) Check whether frontend requested a notification

27.4.1) If so, Perform hypercall `EVTCHNOP_send` which might mean a __VMEXIT__
      depending on the guest type.

**Frontend**

28) Transmit interrupt is raised which signals the packet transmission 
completion.

29) Transmit completion routine checks for unconsumed responses 

30) Processes the responses and revokes the grants provided.

31) Updates `rsp_cons` (request consumer index)

This proposal aims at removing steps 19) 20) 21) by using grefs previously
mapped at guest request. Guest decides how to distribute or use these premapped
grefs with either linear or full packet. This allows us to replace step 27)
(the unmap) preventing the TLB flush.

Note that a grant copy does the following (in pseudo code):

        rcu_lock(src_domain);
        rcu_lock(dst_domain);

        for (op = gntcopy[0]; op < nr_ops; op++) {
                src_frame = __acquire_grant_for_copy(src_domain, <op.src.gref>);
                ^ here implies a holding a potential contended per CPU lock on 
the
                  remote grant table.
                src_vaddr = map_domain_page(src_frame);
        
                dst_frame = __get_paged_frame(dst_domain, <op.dst.mfn>)
                dst_vaddr = map_domain_page(dst_frame);

                memcpy(dst_vaddr + <op.dst.offset>,
                        src_frame + <op.src.offset>,
                        <op.size>);

                unmap_domain_page(src_frame);
                unmap_domain_page(dst_frame);

        rcu_unlock(src_domain);
        rcu_unlock(dst_domain);

Linux netback implementation copies the first 128 bytes into its network buffer
linear region. Hence on the case of the first region it is replaced by a memcpy
on backend, as opposed to a grant copy.

\clearpage

## Guest Receive

The view of the shared receive ring is the following:

     0     1     2     3     4     5     6     7 octet
    +------------------------+------------------------+
    | req_prod               | req_event              |
    +------------------------+------------------------+
    | rsp_prod               | rsp_event              |
    +------------------------+------------------------+
    | pvt                    | pad[44]                |
    +------------------------+                        |
    | ....                                            | [64bytes]
    +------------------------+------------------------+
    | id         | pad       | gref                   | ->'struct 
netif_rx_request'
    +------------+-----------+------------------------+
    | id         | offset    | flags     | status     | ->'struct 
netif_rx_response'
    +-------------------------------------------------+
    |/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/| .. N
    +-------------------------------------------------+

 
Each entry in the ring occupies 16 octets which means a page fits 256 entries.
Additionally a `struct netif_extra_info` may overlay the rx request in which
case the format is:

    +------------------------+------------------------+
    | type |flags| type specific data (gso, hash, etc)| ->'struct 
netif_extra_info'
    +------------+-----------+------------------------+ 

Notice the lack of padding, and that is because it's not used on Rx, as Rx
request boundary is 8 octets.

In essence the steps for receiving of a packet in a Linux frontend is as
 from backend to frontend network stack:

**Backend**

1) Backend transmit function starts

[*Linux-specific*: It means we take a packet and add to an internal queue
 (protected by a lock) whereas a separate thread takes it from that queue and
 process the actual like the steps below. This thread has the purpose of
 aggregating as much copies as possible.]

2) Checks if there are enough rx ring slots that can accomodate the packet.

3) Gets a request from the ring for the first data slot and fetches the `gref`
   from it.

4) Create grant copy op from packet page to `gref`.

[ It's up to the backend to choose how it fills this data. E.g. backend may
  choose to merge as much as data from different pages into this single gref,
  similar to mergeable rx buffers in vhost. ]

5) Sets up flags/checksum info on first request.

6) Gets a response from the ring for this data slot.

7) Prefill expected response ring with the request `id` and slot size.

8) Update the request consumer index (`req_cons`)

9) Gets a request from the ring for the first extra info [optional]

10) Sets up extra info (e.g. GSO descriptor) [optional] repeat step 8).

11) Repeat steps 3 through 8 for all packet pages and set `NETRXF_more_data`
   in the N-1 slot.

12) Perform the `GNTTABOP_copy` hypercall.

13) Check if the grant operations status was incorrect and if so set `status`
    of the `struct netif_rx_response` field to NETIF_RSP_ERR. 

14) Update the response producer index (`rsp_prod`)

**Frontend**

15) Frontend gets an interrupt and runs its interrupt service routine

16) Checks if there's unconsumed responses

17) Consumes a response from the ring (first response for a packet)

18) Revoke the `gref` in the response

19) Consumes extra info response [optional]

20) While N-1 requests has `NETRXF_more_data`, then fetch each of responses
    and revoke the designated `gref`.

21) Update the response consumer index (`rsp_cons`)

22) *Linux-specific*: Copy (from first slot gref) up to 256 bytes to the linear
    region of the packet metadata structure (skb). The rest of the pages
    processed in the responses are then added as frags.

23) Set checksum info based on first response flags.

24) Call packet into the network stack.

25) Allocate new pages and any necessary packet metadata strutures to new
    requests. These requests will then be used in step 1) and so forth.

26) Update the request producer index (`req_prod`)
   
27) Check whether backend needs notification:

27.1) If so, Perform hypercall `EVTCHNOP_send` which might mean a __VMEXIT__
      depending on the guest type.

28) Update `rsp_event` to `response consumer index + 1` such that frontend
    receive a notification on the first newly produced response.
    [optional, if frontend is polling the ring and never sleeps]

This proposal aims at replacing step 4), 12) and  22) with memcpy if the
grefs on the Rx ring were requested to be mapped by the guest. Frontend may use
strategies to allow fast recycling of grants for replinishing the ring,
hence letting Domain-0 replace the grant copies with  memcpy instead, which is
faster.

Depending on the implementation, it would mean that we no longer
would need to aggregate as much as grant ops as possible (step 1) and could
transmit the packet on the transmit function (e.g. Linux ```ndo_start_xmit```)
as previously proposed
here\[[0](http://lists.xenproject.org/archives/html/xen-devel/2015-05/msg01504.html)\].
This would heavily improve efficiency specifially for smaller packets. Which in
return would decrease RTT, having data being acknoledged much quicker.

\clearpage

# Proposed Extension

The idea is to allow guest more controllability on how its grants are mapped or
not. Currently there's no control over it for frontends or backends, and latter
cannot make assumptions on the mapping transmit or receive grants, hence we
need frontend to take initiative into managing its own mapping of grants.
Guests may then opportunistically recycle these grants (e.g. Linux) and avoid
resorting to copies which come when using a fixed amount of buffers. Other
frameworks (e.g.  XDP, netmap, DPDK) use a fixed set of buffers which also
makes the case for this extension.

## Terminology

`staging grants` is a term used in this document to refer to the whole concept
of having a set of grants permanently mapped with backend, containing data
staging until completion. Therefore the term should not be confused with a new
kind of grants on the hypervisor.

## Control Ring Messages

### `XEN_NETIF_CTRL_TYPE_GET_GREF_MAPPING_SIZE`

This message is sent by the frontend to fetch the number of grefs that can
be kept mapped in the backend. It only receives the queue as argument, and
data representing amount of free entries in the mapping table.

### `XEN_NETIF_CTRL_TYPE_ADD_GREF_MAPPING`

This is sent by the frontend to map a list of grant references in the backend.
It receives the queue index, the grant containing the list (offset is
implicitly zero) and how many entries in the list. Each entry in this list
has the following format:

            0     1     2     3     4     5     6     7  octet
         +-----+-----+-----+-----+-----+-----+-----+-----+
         | grant ref             |  flags    |  padding  |
         +-----+-----+-----+-----+-----+-----+-----+-----+

         grant rief: grant reference
         flags: flags describing the control operation

The list can have a maximum of 512 entries to be mapped at once.

### `XEN_NETIF_CTRL_TYPE_PUT_GREF_MAPPING`

This is sent by the frontend for backend to unmap a list of grant
references. The arguments are the same as 
`XEN_NETIF_CTRL_TYPE_ADD_GREF_MAPPING`,
including the format of the list. However entries to be specified on the list
can only refer to the ones previously added with
`XEN_NETIF_CTRL_TYPE_ADD_GREF_MAPPING` and additionally these can not be
inflight grant references in ring at the time the user has requested to unmap
them.

## Datapath Changes

Control ring is only available after backend state is `XenbusConnected`
therefore only on this state change can the frontend query the total amount of
maps it can keep. It then grants N entries per queue on both TX and RX ring
which will create the underying backend gref -> page association (e.g.  stored
in hash table). Frontend may wish to recycle these pregranted buffers or choose
a copy approach to replace granting.

On steps 19) of Guest Transmit and 3) of Guest Receive, data gref is first
looked up in this table and uses the underlying page if it already exists a
mapping. On the successfull cases, steps 20) 21) and 27) of Guest Transmit are
skipped, with 19) being replaced with a memcpy of up to 128 bytes. On Guest
Receive, 4) 12) and 22) are replaced with memcpy instead of a grant copy.

Failing to obtain the total number of mappings
(`XEN_NETIF_CTRL_TYPE_GET_GREF_MAPPING_SIZE`) means the guest falls back to the
normal usage without pre granting buffers.

\clearpage

# Wire Performance

This section is a glossary meant to keep in mind numbers on the wire.

The minimum size that can fit in a single packet with size N is calculated as:

  Packet = Ethernet Header (14) + Protocol Data Unit (46 - 1500) = 60 bytes

In the wire it's a bit more:

  Preamble (7) + Start Frame Delimiter (1) + Packet + CRC (4) + Interframe gap 
(12) = 84 bytes

For given Link-speed in Bits/sec and Packet size, real packet rate is
        calculated as:

  Rate = Link-speed / ((Preamble + Packet + CRC + Interframe gap) * 8)

Numbers to keep in mind (packet size excludes PHY layer, though packet rates
disclosed by vendors take those into account, since it's what goes on the
wire):

| Packet + CRC (bytes)   | 10 Gbit/s  |  40 Gbit/s |  100 Gbit/s  |
|------------------------|:----------:|:----------:|:------------:|
| 64                     | 14.88  Mpps|  59.52 Mpps|  148.80 Mpps |
| 128                    |  8.44  Mpps|  33.78 Mpps|   84.46 Mpps |
| 256                    |  4.52  Mpps|  18.11 Mpps|   45.29 Mpps |
| 1500                   |   822  Kpps|   3.28 Mpps|    8.22 Mpps |
| 65535                  |   ~19  Kpps|  76.27 Kpps|  190.68 Kpps |

Caption:  Mpps (Million packets per second) ; Kpps (Kilo packets per second)

\clearpage

# Performance

Numbers between a Linux v4.11 guest and another host connected by a 100 Gbit/s
Mellanox NIC on a E5-2630 v4 2.2 GHz host to give an idea on the performance
benefits of this extension. Please refer to this presentation[7] for a better
overview of the results.

( Numbers include protocol overhead )

**bulk transfer (Guest TX/RX)**

 Queues  Before (Mbit/s) After (Mbit/s)
 ------  -------------   ------------
 1queue  17244/6000      38189/28108
 2queue  24023/9416      54783/40624
 3queue  29148/17196     85777/54118
 4queue  39782/18502     99530/46859

( Guest -> Dom0 )

**Packet I/O (Guest TX/RX) in UDP 64b**

 Queues  Before (Mpps)  After (Mpps)
 ------  -------------  ------------
 1queue  0.68/0.43      2.49/2.96
 2queue  0.95/0.75      4.74/5.07
 4queue  1.89/1.39      8.80/9.92

\clearpage

# References

[0] http://lists.xenproject.org/archives/html/xen-devel/2015-05/msg01504.html

[1] 
https://github.com/freebsd/freebsd/blob/master/sys/dev/netmap/netmap_mem2.c#L362

[2] https://www.freebsd.org/cgi/man.cgi?query=vale&sektion=4&n=1

[3] https://github.com/iovisor/bpf-docs/blob/master/Express_Data_Path.pdf 

[4] 
http://prototype-kernel.readthedocs.io/en/latest/networking/XDP/design/requirements.html#write-access-to-packet-data

[5] 
http://lxr.free-electrons.com/source/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c#L2073

[6] 
http://lxr.free-electrons.com/source/drivers/net/ethernet/mellanox/mlx4/en_rx.c#L52

[7] 
https://schd.ws/hosted_files/xendeveloperanddesignsummit2017/e6/ToGrantOrNotToGrant-XDDS2017_v3.pdf

# History

A table of changes to the document, in chronological order.

------------------------------------------------------------------------
Date       Revision Version  Notes
---------- -------- -------- -------------------------------------------
2016-12-14 1        Xen 4.9  Initial version for RFC

2016-08-29 2        Xen 4.10 Trim down the specification

                             Rewrite Proposed Extension to use control ring

                             Adjust "Guest Transmission" and "Guest Receive"
                             to reflect the newly proposed extension.

                             Added some performance numbers from the
                             presentation
------------------------------------------------------------------------


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel