1 .. SPDX-License-Identifier: BSD-3-Clause
2 Copyright(c) 2018-2019 Cisco Systems, Inc.
8 Shared memory packet interface (memif) PMD allows for DPDK and any other client
9 using memif (DPDK, VPP, libmemif) to communicate using shared memory. Memif is
12 The created device transmits packets in a raw format. It can be used with
13 Ethernet mode, IP mode, or Punt/Inject. At this moment, only Ethernet mode is
14 supported in DPDK memif implementation.
16 Memif works in two roles: master and slave. Slave connects to master over an
17 existing socket. It is also a producer of shared memory file and initializes
18 the shared memory. Each interface can be connected to one peer interface
19 at same time. The peer interface is identified by id parameter. Master
20 creates the socket and listens for any slave connection requests. The socket
21 may already exist on the system. Be sure to remove any such sockets, if you
22 are creating a master interface, or you will see an "Address already in use"
23 error. Function ``rte_pmd_memif_remove()``, which removes memif interface,
24 will also remove a listener socket, if it is not being used by any other
27 The method to enable one or more interfaces is to use the
28 ``--vdev=net_memif0`` option on the DPDK application command line. Each
29 ``--vdev=net_memif1`` option given will create an interface named net_memif0,
30 net_memif1, and so on. Memif uses unix domain socket to transmit control
31 messages. Each memif has a unique id per socket. This id is used to identify
32 peer interface. If you are connecting multiple
33 interfaces using same socket, be sure to specify unique ids ``id=0``, ``id=1``,
34 etc. Note that if you assign a socket to a master interface it becomes a
35 listener socket. Listener socket can not be used by a slave interface on same
38 .. csv-table:: **Memif configuration options**
39 :header: "Option", "Description", "Default", "Valid value"
41 "id=0", "Used to identify peer interface", "0", "uint32_t"
42 "role=master", "Set memif role", "slave", "master|slave"
43 "bsize=1024", "Size of single packet buffer", "2048", "uint16_t"
44 "rsize=11", "Log2 of ring size. If rsize is 10, actual ring size is 1024", "10", "1-14"
45 "socket=/tmp/memif.sock", "Socket filename", "/tmp/memif.sock", "string len 108"
46 "mac=01:23:45:ab:cd:ef", "Mac address", "01:ab:23:cd:45:ef", ""
47 "secret=abc123", "Secret is an optional security option, which if specified, must be matched by peer", "", "string len 24"
48 "zero-copy=yes", "Enable/disable zero-copy slave mode. Only relevant to slave, requires '--single-file-segments' eal argument", "no", "yes|no"
50 **Connection establishment**
52 In order to create memif connection, two memif interfaces, each in separate
53 process, are needed. One interface in ``master`` role and other in
54 ``slave`` role. It is not possible to connect two interfaces in a single
55 process. Each interface can be connected to one interface at same time,
56 identified by matching id parameter.
58 Memif driver uses unix domain socket to exchange required information between
59 memif interfaces. Socket file path is specified at interface creation see
60 *Memif configuration options* table above. If socket is used by ``master``
61 interface, it's marked as listener socket (in scope of current process) and
62 listens to connection requests from other processes. One socket can be used by
63 multiple interfaces. One process can have ``slave`` and ``master`` interfaces
64 at the same time, provided each role is assigned unique socket.
66 For detailed information on memif control messages, see: net/memif/memif.h.
68 Slave interface attempts to make a connection on assigned socket. Process
69 listening on this socket will extract the connection request and create a new
70 connected socket (control channel). Then it sends the 'hello' message
71 (``MEMIF_MSG_TYPE_HELLO``), containing configuration boundaries. Slave interface
72 adjusts its configuration accordingly, and sends 'init' message
73 (``MEMIF_MSG_TYPE_INIT``). This message among others contains interface id. Driver
74 uses this id to find master interface, and assigns the control channel to this
75 interface. If such interface is found, 'ack' message (``MEMIF_MSG_TYPE_ACK``) is
76 sent. Slave interface sends 'add region' message (``MEMIF_MSG_TYPE_ADD_REGION``) for
77 every region allocated. Master responds to each of these messages with 'ack'
78 message. Same behavior applies to rings. Slave sends 'add ring' message
79 (``MEMIF_MSG_TYPE_ADD_RING``) for every initialized ring. Master again responds to
80 each message with 'ack' message. To finalize the connection, slave interface
81 sends 'connect' message (``MEMIF_MSG_TYPE_CONNECT``). Upon receiving this message
82 master maps regions to its address space, initializes rings and responds with
83 'connected' message (``MEMIF_MSG_TYPE_CONNECTED``). Disconnect
84 (``MEMIF_MSG_TYPE_DISCONNECT``) can be sent by both master and slave interfaces at
85 any time, due to driver error or if the interface is being deleted.
89 - net/memif/memif.h *- control messages definitions*
90 - net/memif/memif_socket.h
91 - net/memif/memif_socket.c
96 **Shared memory format**
98 Slave is producer and master is consumer. Memory regions, are mapped shared memory files,
99 created by memif slave and provided to master at connection establishment.
100 Regions contain rings and buffers. Rings and buffers can also be separated into multiple
101 regions. For no-zero-copy, rings and buffers are stored inside single memory
102 region to reduce the number of opened files.
104 region n (no-zero-copy):
106 +-----------------------+-------------------------------------------------------------------------+
108 +-----------+-----------+-----------------+---+---------------------------------------------------+
109 | S2M rings | M2S rings | packet buffer 0 | . | pb ((1 << pmd->run.log2_ring_size)*(s2m + m2s))-1 |
110 +-----------+-----------+-----------------+---+---------------------------------------------------+
114 +--------+--------+-----------------------+
115 | ring 0 | ring 1 | ring num_s2m_rings - 1|
116 +--------+--------+-----------------------+
120 +-------------+---------------------------------------+
121 | ring header | (1 << pmd->run.log2_ring_size) * desc |
122 +-------------+---------------------------------------+
124 Descriptors are assigned packet buffers in order of rings creation. If we have one ring
125 in each direction and ring size is 1024, then first 1024 buffers will belong to S2M ring and
126 last 1024 will belong to M2S ring. In case of zero-copy, buffers are dequeued and
129 **Descriptor format**
131 +----+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
132 |Quad|6| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |3|3| | | | | | | | | | | | | | |1|1| | | | | | | | | | | | | | | |
133 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
134 |Word|3| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |2|1| | | | | | | | | | | | | | |6|5| | | | | | | | | | | | | | |0|
135 +----+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
136 |0 |length |region |flags |
137 +----+---------------------------------------------------------------+-------------------------------+-------------------------------+
138 |1 |metadata |offset |
139 +----+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
140 | |6| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |3|3| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
141 | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
142 | |3| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |2|1| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |0|
143 +----+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
145 **Flags field - flags (Quad Word 0, bits 0:15)**
147 +-----+--------------------+------------------------------------------------------------------------------------------------+
148 |Bits |Name |Functionality |
149 +=====+====================+================================================================================================+
150 |0 |MEMIF_DESC_FLAG_NEXT|Is chained buffer. When set, the packet is divided into multiple buffers. May not be contiguous.|
151 +-----+--------------------+------------------------------------------------------------------------------------------------+
153 **Region index - region (Quad Word 0, 16:31)**
155 Index of memory region, the buffer is located in.
157 **Data length - length (Quad Word 0, 32:63)**
159 Length of transmitted/received data.
161 **Data Offset - offset (Quad Word 1, 0:31)**
163 Data start offset from memory region address. *.regions[desc->region].addr + desc->offset*
165 **Metadata - metadata (Quad Word 1, 32:63)**
171 - net/memif/memif.h *- descriptor and ring definitions*
172 - net/memif/rte_eth_memif.c *- eth_memif_rx() eth_memif_tx()*
177 Zero-copy slave can be enabled with memif configuration option 'zero-copy=yes'. This option
178 is only relevant to slave and requires eal argument '--single-file-segments'.
179 This limitation is in place, because it is too expensive to identify memseg
180 for each packet buffer, resulting in worse performance than with zero-copy disabled.
181 With single file segments we can calculate offset from the beginning of the file
182 for each packet buffer.
184 **Shared memory format**
186 Region 0 is created by memif driver and contains rings. Slave interface exposes DPDK memory (memseg).
187 Instead of using memfd_create() to create new shared file, existing memsegs are used.
188 Master interface functions the same as with zero-copy disabled.
192 +-----------------------+
194 +-----------+-----------+
195 | S2M rings | M2S rings |
196 +-----------+-----------+
206 Buffers are dequeued and enqueued as needed. Offset descriptor field is calculated at tx.
207 Only single file segments mode (EAL option --single-file-segments) is supported, as calculating
208 offset from multiple segments is too expensive.
211 ----------------------------
212 In this example we run two instances of testpmd application and transmit packets over memif.
214 First create ``master`` interface::
216 #./build/app/testpmd -l 0-1 --proc-type=primary --file-prefix=pmd1 --vdev=net_memif,role=master -- -i
218 Now create ``slave`` interface (master must be already running so the slave will connect)::
220 #./build/app/testpmd -l 2-3 --proc-type=primary --file-prefix=pmd2 --vdev=net_memif -- -i
222 You can also enable ``zero-copy`` on ``slave`` interface::
224 #./build/app/testpmd -l 2-3 --proc-type=primary --file-prefix=pmd2 --vdev=net_memif,zero-copy=yes --single-file-segments -- -i
226 Start forwarding packets::
232 testpmd> start tx_first
236 testpmd> show port stats 0
238 For more details on testpmd please refer to :doc:`../testpmd_app_ug/index`.
240 Example: testpmd and VPP
241 ------------------------
242 For information on how to get and run VPP please see `<https://wiki.fd.io/view/VPP>`_.
244 Start VPP in interactive mode (should be by default). Create memif master interface in VPP::
246 vpp# create interface memif id 0 master no-zero-copy
247 vpp# set interface state memif0/0 up
248 vpp# set interface ip address memif0/0 192.168.1.1/24
250 To see socket filename use show memif command::
255 0 yes (1) /run/vpp/memif.sock
258 Now create memif interface by running testpmd with these command line options::
260 #./testpmd --vdev=net_memif,socket=/run/vpp/memif.sock -- -i
262 Testpmd should now create memif slave interface and try to connect to master.
263 In testpmd set forward option to icmpecho and start forwarding::
265 testpmd> set fwd icmpecho
270 vpp# ping 192.168.1.2
271 64 bytes from 192.168.1.2: icmp_seq=2 ttl=254 time=36.2918 ms
272 64 bytes from 192.168.1.2: icmp_seq=3 ttl=254 time=23.3927 ms
273 64 bytes from 192.168.1.2: icmp_seq=4 ttl=254 time=24.2975 ms
274 64 bytes from 192.168.1.2: icmp_seq=5 ttl=254 time=17.7049 ms