1.. SPDX-License-Identifier: GPL-2.0
2
3===========
4Packet MMAP
5===========
6
7Abstract
8========
9
10This file documents the mmap() facility available with the PACKET
11socket interface. This type of sockets is used for
12
13i) capture network traffic with utilities like tcpdump,
14ii) transmit network traffic, or any other that needs raw
15    access to network interface.
16
17Howto can be found at:
18
19    https://sites.google.com/site/packetmmap/
20
21Please send your comments to
22    - Ulisses Alonso Camaró <uaca@i.hate.spam.alumni.uv.es>
23    - Johann Baudy
24
25Why use PACKET_MMAP
26===================
27
28Non PACKET_MMAP capture process (plain AF_PACKET) is very
29inefficient. It uses very limited buffers and requires one system call to
30capture each packet, it requires two if you want to get packet's timestamp
31(like libpcap always does).
32
33On the other hand PACKET_MMAP is very efficient. PACKET_MMAP provides a size
34configurable circular buffer mapped in user space that can be used to either
35send or receive packets. This way reading packets just needs to wait for them,
36most of the time there is no need to issue a single system call. Concerning
37transmission, multiple packets can be sent through one system call to get the
38highest bandwidth. By using a shared buffer between the kernel and the user
39also has the benefit of minimizing packet copies.
40
41It's fine to use PACKET_MMAP to improve the performance of the capture and
42transmission process, but it isn't everything. At least, if you are capturing
43at high speeds (this is relative to the cpu speed), you should check if the
44device driver of your network interface card supports some sort of interrupt
45load mitigation or (even better) if it supports NAPI, also make sure it is
46enabled. For transmission, check the MTU (Maximum Transmission Unit) used and
47supported by devices of your network. CPU IRQ pinning of your network interface
48card can also be an advantage.
49
50How to use mmap() to improve capture process
51============================================
52
53From the user standpoint, you should use the higher level libpcap library, which
54is a de facto standard, portable across nearly all operating systems
55including Win32.
56
57Packet MMAP support was integrated into libpcap around the time of version 1.3.0;
58TPACKET_V3 support was added in version 1.5.0
59
60How to use mmap() directly to improve capture process
61=====================================================
62
63From the system calls stand point, the use of PACKET_MMAP involves
64the following process::
65
66
67    [setup]     socket() -------> creation of the capture socket
68		setsockopt() ---> allocation of the circular buffer (ring)
69				  option: PACKET_RX_RING
70		mmap() ---------> mapping of the allocated buffer to the
71				  user process
72
73    [capture]   poll() ---------> to wait for incoming packets
74
75    [shutdown]  close() --------> destruction of the capture socket and
76				  deallocation of all associated
77				  resources.
78
79
80socket creation and destruction is straight forward, and is done
81the same way with or without PACKET_MMAP::
82
83 int fd = socket(PF_PACKET, mode, htons(ETH_P_ALL));
84
85where mode is SOCK_RAW for the raw interface were link level
86information can be captured or SOCK_DGRAM for the cooked
87interface where link level information capture is not
88supported and a link level pseudo-header is provided
89by the kernel.
90
91The destruction of the socket and all associated resources
92is done by a simple call to close(fd).
93
94Similarly as without PACKET_MMAP, it is possible to use one socket
95for capture and transmission. This can be done by mapping the
96allocated RX and TX buffer ring with a single mmap() call.
97See "Mapping and use of the circular buffer (ring)".
98
99Next I will describe PACKET_MMAP settings and its constraints,
100also the mapping of the circular buffer in the user process and
101the use of this buffer.
102
103How to use mmap() directly to improve transmission process
104==========================================================
105Transmission process is similar to capture as shown below::
106
107    [setup]         socket() -------> creation of the transmission socket
108		    setsockopt() ---> allocation of the circular buffer (ring)
109				      option: PACKET_TX_RING
110		    bind() ---------> bind transmission socket with a network interface
111		    mmap() ---------> mapping of the allocated buffer to the
112				      user process
113
114    [transmission]  poll() ---------> wait for free packets (optional)
115		    send() ---------> send all packets that are set as ready in
116				      the ring
117				      The flag MSG_DONTWAIT can be used to return
118				      before end of transfer.
119
120    [shutdown]      close() --------> destruction of the transmission socket and
121				      deallocation of all associated resources.
122
123Socket creation and destruction is also straight forward, and is done
124the same way as in capturing described in the previous paragraph::
125
126 int fd = socket(PF_PACKET, mode, 0);
127
128The protocol can optionally be 0 in case we only want to transmit
129via this socket, which avoids an expensive call to packet_rcv().
130In this case, you also need to bind(2) the TX_RING with sll_protocol = 0
131set. Otherwise, htons(ETH_P_ALL) or any other protocol, for example.
132
133Binding the socket to your network interface is mandatory (with zero copy) to
134know the header size of frames used in the circular buffer.
135
136As capture, each frame contains two parts::
137
138    --------------------
139    | struct tpacket_hdr | Header. It contains the status of
140    |                    | of this frame
141    |--------------------|
142    | data buffer        |
143    .                    .  Data that will be sent over the network interface.
144    .                    .
145    --------------------
146
147 bind() associates the socket to your network interface thanks to
148 sll_ifindex parameter of struct sockaddr_ll.
149
150 Initialization example::
151
152    struct sockaddr_ll my_addr;
153    struct ifreq s_ifr;
154    ...
155
156    strncpy (s_ifr.ifr_name, "eth0", sizeof(s_ifr.ifr_name));
157
158    /* get interface index of eth0 */
159    ioctl(this->socket, SIOCGIFINDEX, &s_ifr);
160
161    /* fill sockaddr_ll struct to prepare binding */
162    my_addr.sll_family = AF_PACKET;
163    my_addr.sll_protocol = htons(ETH_P_ALL);
164    my_addr.sll_ifindex =  s_ifr.ifr_ifindex;
165
166    /* bind socket to eth0 */
167    bind(this->socket, (struct sockaddr *)&my_addr, sizeof(struct sockaddr_ll));
168
169 A complete tutorial is available at: https://sites.google.com/site/packetmmap/
170
171By default, the user should put data at::
172
173 frame base + TPACKET_HDRLEN - sizeof(struct sockaddr_ll)
174
175So, whatever you choose for the socket mode (SOCK_DGRAM or SOCK_RAW),
176the beginning of the user data will be at::
177
178 frame base + TPACKET_ALIGN(sizeof(struct tpacket_hdr))
179
180If you wish to put user data at a custom offset from the beginning of
181the frame (for payload alignment with SOCK_RAW mode for instance) you
182can set tp_net (with SOCK_DGRAM) or tp_mac (with SOCK_RAW). In order
183to make this work it must be enabled previously with setsockopt()
184and the PACKET_TX_HAS_OFF option.
185
186PACKET_MMAP settings
187====================
188
189To setup PACKET_MMAP from user level code is done with a call like
190
191 - Capture process::
192
193     setsockopt(fd, SOL_PACKET, PACKET_RX_RING, (void *) &req, sizeof(req))
194
195 - Transmission process::
196
197     setsockopt(fd, SOL_PACKET, PACKET_TX_RING, (void *) &req, sizeof(req))
198
199The most significant argument in the previous call is the req parameter,
200this parameter must to have the following structure::
201
202    struct tpacket_req
203    {
204	unsigned int    tp_block_size;  /* Minimal size of contiguous block */
205	unsigned int    tp_block_nr;    /* Number of blocks */
206	unsigned int    tp_frame_size;  /* Size of frame */
207	unsigned int    tp_frame_nr;    /* Total number of frames */
208    };
209
210This structure is defined in /usr/include/linux/if_packet.h and establishes a
211circular buffer (ring) of unswappable memory.
212Being mapped in the capture process allows reading the captured frames and
213related meta-information like timestamps without requiring a system call.
214
215Frames are grouped in blocks. Each block is a physically contiguous
216region of memory and holds tp_block_size/tp_frame_size frames. The total number
217of blocks is tp_block_nr. Note that tp_frame_nr is a redundant parameter because::
218
219    frames_per_block = tp_block_size/tp_frame_size
220
221indeed, packet_set_ring checks that the following condition is true::
222
223    frames_per_block * tp_block_nr == tp_frame_nr
224
225Lets see an example, with the following values::
226
227     tp_block_size= 4096
228     tp_frame_size= 2048
229     tp_block_nr  = 4
230     tp_frame_nr  = 8
231
232we will get the following buffer structure::
233
234	    block #1                 block #2
235    +---------+---------+    +---------+---------+
236    | frame 1 | frame 2 |    | frame 3 | frame 4 |
237    +---------+---------+    +---------+---------+
238
239	    block #3                 block #4
240    +---------+---------+    +---------+---------+
241    | frame 5 | frame 6 |    | frame 7 | frame 8 |
242    +---------+---------+    +---------+---------+
243
244A frame can be of any size with the only condition it can fit in a block. A block
245can only hold an integer number of frames, or in other words, a frame cannot
246be spawned across two blocks, so there are some details you have to take into
247account when choosing the frame_size. See "Mapping and use of the circular
248buffer (ring)".
249
250PACKET_MMAP setting constraints
251===============================
252
253In kernel versions prior to 2.4.26 (for the 2.4 branch) and 2.6.5 (2.6 branch),
254the PACKET_MMAP buffer could hold only 32768 frames in a 32 bit architecture or
25516384 in a 64 bit architecture.
256
257Block size limit
258----------------
259
260As stated earlier, each block is a contiguous physical region of memory. These
261memory regions are allocated with calls to the __get_free_pages() function. As
262the name indicates, this function allocates pages of memory, and the second
263argument is "order" or a power of two number of pages, that is
264(for PAGE_SIZE == 4096) order=0 ==> 4096 bytes, order=1 ==> 8192 bytes,
265order=2 ==> 16384 bytes, etc. The maximum size of a
266region allocated by __get_free_pages is determined by the MAX_ORDER macro. More
267precisely the limit can be calculated as::
268
269   PAGE_SIZE << MAX_ORDER
270
271   In a i386 architecture PAGE_SIZE is 4096 bytes
272   In a 2.4/i386 kernel MAX_ORDER is 10
273   In a 2.6/i386 kernel MAX_ORDER is 11
274
275So get_free_pages can allocate as much as 4MB or 8MB in a 2.4/2.6 kernel
276respectively, with an i386 architecture.
277
278User space programs can include /usr/include/sys/user.h and
279/usr/include/linux/mmzone.h to get PAGE_SIZE MAX_ORDER declarations.
280
281The pagesize can also be determined dynamically with the getpagesize (2)
282system call.
283
284Block number limit
285------------------
286
287To understand the constraints of PACKET_MMAP, we have to see the structure
288used to hold the pointers to each block.
289
290Currently, this structure is a dynamically allocated vector with kmalloc
291called pg_vec, its size limits the number of blocks that can be allocated::
292
293    +---+---+---+---+
294    | x | x | x | x |
295    +---+---+---+---+
296      |   |   |   |
297      |   |   |   v
298      |   |   v  block #4
299      |   v  block #3
300      v  block #2
301     block #1
302
303kmalloc allocates any number of bytes of physically contiguous memory from
304a pool of pre-determined sizes. This pool of memory is maintained by the slab
305allocator which is at the end the responsible for doing the allocation and
306hence which imposes the maximum memory that kmalloc can allocate.
307
308In a 2.4/2.6 kernel and the i386 architecture, the limit is 131072 bytes. The
309predetermined sizes that kmalloc uses can be checked in the "size-<bytes>"
310entries of /proc/slabinfo
311
312In a 32 bit architecture, pointers are 4 bytes long, so the total number of
313pointers to blocks is::
314
315     131072/4 = 32768 blocks
316
317PACKET_MMAP buffer size calculator
318==================================
319
320Definitions:
321
322==============  ================================================================
323<size-max>      is the maximum size of allocable with kmalloc
324		(see /proc/slabinfo)
325<pointer size>  depends on the architecture -- ``sizeof(void *)``
326<page size>     depends on the architecture -- PAGE_SIZE or getpagesize (2)
327<max-order>     is the value defined with MAX_ORDER
328<frame size>    it's an upper bound of frame's capture size (more on this later)
329==============  ================================================================
330
331from these definitions we will derive::
332
333	<block number> = <size-max>/<pointer size>
334	<block size> = <pagesize> << <max-order>
335
336so, the max buffer size is::
337
338	<block number> * <block size>
339
340and, the number of frames be::
341
342	<block number> * <block size> / <frame size>
343
344Suppose the following parameters, which apply for 2.6 kernel and an
345i386 architecture::
346
347	<size-max> = 131072 bytes
348	<pointer size> = 4 bytes
349	<pagesize> = 4096 bytes
350	<max-order> = 11
351
352and a value for <frame size> of 2048 bytes. These parameters will yield::
353
354	<block number> = 131072/4 = 32768 blocks
355	<block size> = 4096 << 11 = 8 MiB.
356
357and hence the buffer will have a 262144 MiB size. So it can hold
358262144 MiB / 2048 bytes = 134217728 frames
359
360Actually, this buffer size is not possible with an i386 architecture.
361Remember that the memory is allocated in kernel space, in the case of
362an i386 kernel's memory size is limited to 1GiB.
363
364All memory allocations are not freed until the socket is closed. The memory
365allocations are done with GFP_KERNEL priority, this basically means that
366the allocation can wait and swap other process' memory in order to allocate
367the necessary memory, so normally limits can be reached.
368
369Other constraints
370-----------------
371
372If you check the source code you will see that what I draw here as a frame
373is not only the link level frame. At the beginning of each frame there is a
374header called struct tpacket_hdr used in PACKET_MMAP to hold link level's frame
375meta information like timestamp. So what we draw here a frame it's really
376the following (from include/linux/if_packet.h)::
377
378 /*
379   Frame structure:
380
381   - Start. Frame must be aligned to TPACKET_ALIGNMENT=16
382   - struct tpacket_hdr
383   - pad to TPACKET_ALIGNMENT=16
384   - struct sockaddr_ll
385   - Gap, chosen so that packet data (Start+tp_net) aligns to
386     TPACKET_ALIGNMENT=16
387   - Start+tp_mac: [ Optional MAC header ]
388   - Start+tp_net: Packet data, aligned to TPACKET_ALIGNMENT=16.
389   - Pad to align to TPACKET_ALIGNMENT=16
390 */
391
392The following are conditions that are checked in packet_set_ring
393
394   - tp_block_size must be a multiple of PAGE_SIZE (1)
395   - tp_frame_size must be greater than TPACKET_HDRLEN (obvious)
396   - tp_frame_size must be a multiple of TPACKET_ALIGNMENT
397   - tp_frame_nr   must be exactly frames_per_block*tp_block_nr
398
399Note that tp_block_size should be chosen to be a power of two or there will
400be a waste of memory.
401
402Mapping and use of the circular buffer (ring)
403---------------------------------------------
404
405The mapping of the buffer in the user process is done with the conventional
406mmap function. Even the circular buffer is compound of several physically
407discontiguous blocks of memory, they are contiguous to the user space, hence
408just one call to mmap is needed::
409
410    mmap(0, size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
411
412If tp_frame_size is a divisor of tp_block_size frames will be
413contiguously spaced by tp_frame_size bytes. If not, each
414tp_block_size/tp_frame_size frames there will be a gap between
415the frames. This is because a frame cannot be spawn across two
416blocks.
417
418To use one socket for capture and transmission, the mapping of both the
419RX and TX buffer ring has to be done with one call to mmap::
420
421    ...
422    setsockopt(fd, SOL_PACKET, PACKET_RX_RING, &foo, sizeof(foo));
423    setsockopt(fd, SOL_PACKET, PACKET_TX_RING, &bar, sizeof(bar));
424    ...
425    rx_ring = mmap(0, size * 2, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
426    tx_ring = rx_ring + size;
427
428RX must be the first as the kernel maps the TX ring memory right
429after the RX one.
430
431At the beginning of each frame there is an status field (see
432struct tpacket_hdr). If this field is 0 means that the frame is ready
433to be used for the kernel, If not, there is a frame the user can read
434and the following flags apply:
435
436Capture process
437^^^^^^^^^^^^^^^
438
439From include/linux/if_packet.h::
440
441     #define TP_STATUS_COPY          (1 << 1)
442     #define TP_STATUS_LOSING        (1 << 2)
443     #define TP_STATUS_CSUMNOTREADY  (1 << 3)
444     #define TP_STATUS_CSUM_VALID    (1 << 7)
445
446======================  =======================================================
447TP_STATUS_COPY		This flag indicates that the frame (and associated
448			meta information) has been truncated because it's
449			larger than tp_frame_size. This packet can be
450			read entirely with recvfrom().
451
452			In order to make this work it must to be
453			enabled previously with setsockopt() and
454			the PACKET_COPY_THRESH option.
455
456			The number of frames that can be buffered to
457			be read with recvfrom is limited like a normal socket.
458			See the SO_RCVBUF option in the socket (7) man page.
459
460TP_STATUS_LOSING	indicates there were packet drops from last time
461			statistics where checked with getsockopt() and
462			the PACKET_STATISTICS option.
463
464TP_STATUS_CSUMNOTREADY	currently it's used for outgoing IP packets which
465			its checksum will be done in hardware. So while
466			reading the packet we should not try to check the
467			checksum.
468
469TP_STATUS_CSUM_VALID	This flag indicates that at least the transport
470			header checksum of the packet has been already
471			validated on the kernel side. If the flag is not set
472			then we are free to check the checksum by ourselves
473			provided that TP_STATUS_CSUMNOTREADY is also not set.
474======================  =======================================================
475
476for convenience there are also the following defines::
477
478     #define TP_STATUS_KERNEL        0
479     #define TP_STATUS_USER          1
480
481The kernel initializes all frames to TP_STATUS_KERNEL, when the kernel
482receives a packet it puts in the buffer and updates the status with
483at least the TP_STATUS_USER flag. Then the user can read the packet,
484once the packet is read the user must zero the status field, so the kernel
485can use again that frame buffer.
486
487The user can use poll (any other variant should apply too) to check if new
488packets are in the ring::
489
490    struct pollfd pfd;
491
492    pfd.fd = fd;
493    pfd.revents = 0;
494    pfd.events = POLLIN|POLLRDNORM|POLLERR;
495
496    if (status == TP_STATUS_KERNEL)
497	retval = poll(&pfd, 1, timeout);
498
499It doesn't incur in a race condition to first check the status value and
500then poll for frames.
501
502Transmission process
503^^^^^^^^^^^^^^^^^^^^
504
505Those defines are also used for transmission::
506
507     #define TP_STATUS_AVAILABLE        0 // Frame is available
508     #define TP_STATUS_SEND_REQUEST     1 // Frame will be sent on next send()
509     #define TP_STATUS_SENDING          2 // Frame is currently in transmission
510     #define TP_STATUS_WRONG_FORMAT     4 // Frame format is not correct
511
512First, the kernel initializes all frames to TP_STATUS_AVAILABLE. To send a
513packet, the user fills a data buffer of an available frame, sets tp_len to
514current data buffer size and sets its status field to TP_STATUS_SEND_REQUEST.
515This can be done on multiple frames. Once the user is ready to transmit, it
516calls send(). Then all buffers with status equal to TP_STATUS_SEND_REQUEST are
517forwarded to the network device. The kernel updates each status of sent
518frames with TP_STATUS_SENDING until the end of transfer.
519
520At the end of each transfer, buffer status returns to TP_STATUS_AVAILABLE.
521
522::
523
524    header->tp_len = in_i_size;
525    header->tp_status = TP_STATUS_SEND_REQUEST;
526    retval = send(this->socket, NULL, 0, 0);
527
528The user can also use poll() to check if a buffer is available:
529
530(status == TP_STATUS_SENDING)
531
532::
533
534    struct pollfd pfd;
535    pfd.fd = fd;
536    pfd.revents = 0;
537    pfd.events = POLLOUT;
538    retval = poll(&pfd, 1, timeout);
539
540What TPACKET versions are available and when to use them?
541=========================================================
542
543::
544
545 int val = tpacket_version;
546 setsockopt(fd, SOL_PACKET, PACKET_VERSION, &val, sizeof(val));
547 getsockopt(fd, SOL_PACKET, PACKET_VERSION, &val, sizeof(val));
548
549where 'tpacket_version' can be TPACKET_V1 (default), TPACKET_V2, TPACKET_V3.
550
551TPACKET_V1:
552	- Default if not otherwise specified by setsockopt(2)
553	- RX_RING, TX_RING available
554
555TPACKET_V1 --> TPACKET_V2:
556	- Made 64 bit clean due to unsigned long usage in TPACKET_V1
557	  structures, thus this also works on 64 bit kernel with 32 bit
558	  userspace and the like
559	- Timestamp resolution in nanoseconds instead of microseconds
560	- RX_RING, TX_RING available
561	- VLAN metadata information available for packets
562	  (TP_STATUS_VLAN_VALID, TP_STATUS_VLAN_TPID_VALID),
563	  in the tpacket2_hdr structure:
564
565		- TP_STATUS_VLAN_VALID bit being set into the tp_status field indicates
566		  that the tp_vlan_tci field has valid VLAN TCI value
567		- TP_STATUS_VLAN_TPID_VALID bit being set into the tp_status field
568		  indicates that the tp_vlan_tpid field has valid VLAN TPID value
569
570	- How to switch to TPACKET_V2:
571
572		1. Replace struct tpacket_hdr by struct tpacket2_hdr
573		2. Query header len and save
574		3. Set protocol version to 2, set up ring as usual
575		4. For getting the sockaddr_ll,
576		   use ``(void *)hdr + TPACKET_ALIGN(hdrlen)`` instead of
577		   ``(void *)hdr + TPACKET_ALIGN(sizeof(struct tpacket_hdr))``
578
579TPACKET_V2 --> TPACKET_V3:
580	- Flexible buffer implementation for RX_RING:
581		1. Blocks can be configured with non-static frame-size
582		2. Read/poll is at a block-level (as opposed to packet-level)
583		3. Added poll timeout to avoid indefinite user-space wait
584		   on idle links
585		4. Added user-configurable knobs:
586
587			4.1 block::timeout
588			4.2 tpkt_hdr::sk_rxhash
589
590	- RX Hash data available in user space
591	- TX_RING semantics are conceptually similar to TPACKET_V2;
592	  use tpacket3_hdr instead of tpacket2_hdr, and TPACKET3_HDRLEN
593	  instead of TPACKET2_HDRLEN. In the current implementation,
594	  the tp_next_offset field in the tpacket3_hdr MUST be set to
595	  zero, indicating that the ring does not hold variable sized frames.
596	  Packets with non-zero values of tp_next_offset will be dropped.
597
598AF_PACKET fanout mode
599=====================
600
601In the AF_PACKET fanout mode, packet reception can be load balanced among
602processes. This also works in combination with mmap(2) on packet sockets.
603
604Currently implemented fanout policies are:
605
606  - PACKET_FANOUT_HASH: schedule to socket by skb's packet hash
607  - PACKET_FANOUT_LB: schedule to socket by round-robin
608  - PACKET_FANOUT_CPU: schedule to socket by CPU packet arrives on
609  - PACKET_FANOUT_RND: schedule to socket by random selection
610  - PACKET_FANOUT_ROLLOVER: if one socket is full, rollover to another
611  - PACKET_FANOUT_QM: schedule to socket by skbs recorded queue_mapping
612
613Minimal example code by David S. Miller (try things like "./test eth0 hash",
614"./test eth0 lb", etc.)::
615
616    #include <stddef.h>
617    #include <stdlib.h>
618    #include <stdio.h>
619    #include <string.h>
620
621    #include <sys/types.h>
622    #include <sys/wait.h>
623    #include <sys/socket.h>
624    #include <sys/ioctl.h>
625
626    #include <unistd.h>
627
628    #include <linux/if_ether.h>
629    #include <linux/if_packet.h>
630
631    #include <net/if.h>
632
633    static const char *device_name;
634    static int fanout_type;
635    static int fanout_id;
636
637    #ifndef PACKET_FANOUT
638    # define PACKET_FANOUT			18
639    # define PACKET_FANOUT_HASH		0
640    # define PACKET_FANOUT_LB		1
641    #endif
642
643    static int setup_socket(void)
644    {
645	    int err, fd = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_IP));
646	    struct sockaddr_ll ll;
647	    struct ifreq ifr;
648	    int fanout_arg;
649
650	    if (fd < 0) {
651		    perror("socket");
652		    return EXIT_FAILURE;
653	    }
654
655	    memset(&ifr, 0, sizeof(ifr));
656	    strcpy(ifr.ifr_name, device_name);
657	    err = ioctl(fd, SIOCGIFINDEX, &ifr);
658	    if (err < 0) {
659		    perror("SIOCGIFINDEX");
660		    return EXIT_FAILURE;
661	    }
662
663	    memset(&ll, 0, sizeof(ll));
664	    ll.sll_family = AF_PACKET;
665	    ll.sll_ifindex = ifr.ifr_ifindex;
666	    err = bind(fd, (struct sockaddr *) &ll, sizeof(ll));
667	    if (err < 0) {
668		    perror("bind");
669		    return EXIT_FAILURE;
670	    }
671
672	    fanout_arg = (fanout_id | (fanout_type << 16));
673	    err = setsockopt(fd, SOL_PACKET, PACKET_FANOUT,
674			    &fanout_arg, sizeof(fanout_arg));
675	    if (err) {
676		    perror("setsockopt");
677		    return EXIT_FAILURE;
678	    }
679
680	    return fd;
681    }
682
683    static void fanout_thread(void)
684    {
685	    int fd = setup_socket();
686	    int limit = 10000;
687
688	    if (fd < 0)
689		    exit(fd);
690
691	    while (limit-- > 0) {
692		    char buf[1600];
693		    int err;
694
695		    err = read(fd, buf, sizeof(buf));
696		    if (err < 0) {
697			    perror("read");
698			    exit(EXIT_FAILURE);
699		    }
700		    if ((limit % 10) == 0)
701			    fprintf(stdout, "(%d) \n", getpid());
702	    }
703
704	    fprintf(stdout, "%d: Received 10000 packets\n", getpid());
705
706	    close(fd);
707	    exit(0);
708    }
709
710    int main(int argc, char **argp)
711    {
712	    int fd, err;
713	    int i;
714
715	    if (argc != 3) {
716		    fprintf(stderr, "Usage: %s INTERFACE {hash|lb}\n", argp[0]);
717		    return EXIT_FAILURE;
718	    }
719
720	    if (!strcmp(argp[2], "hash"))
721		    fanout_type = PACKET_FANOUT_HASH;
722	    else if (!strcmp(argp[2], "lb"))
723		    fanout_type = PACKET_FANOUT_LB;
724	    else {
725		    fprintf(stderr, "Unknown fanout type [%s]\n", argp[2]);
726		    exit(EXIT_FAILURE);
727	    }
728
729	    device_name = argp[1];
730	    fanout_id = getpid() & 0xffff;
731
732	    for (i = 0; i < 4; i++) {
733		    pid_t pid = fork();
734
735		    switch (pid) {
736		    case 0:
737			    fanout_thread();
738
739		    case -1:
740			    perror("fork");
741			    exit(EXIT_FAILURE);
742		    }
743	    }
744
745	    for (i = 0; i < 4; i++) {
746		    int status;
747
748		    wait(&status);
749	    }
750
751	    return 0;
752    }
753
754AF_PACKET TPACKET_V3 example
755============================
756
757AF_PACKET's TPACKET_V3 ring buffer can be configured to use non-static frame
758sizes by doing it's own memory management. It is based on blocks where polling
759works on a per block basis instead of per ring as in TPACKET_V2 and predecessor.
760
761It is said that TPACKET_V3 brings the following benefits:
762
763 * ~15% - 20% reduction in CPU-usage
764 * ~20% increase in packet capture rate
765 * ~2x increase in packet density
766 * Port aggregation analysis
767 * Non static frame size to capture entire packet payload
768
769So it seems to be a good candidate to be used with packet fanout.
770
771Minimal example code by Daniel Borkmann based on Chetan Loke's lolpcap (compile
772it with gcc -Wall -O2 blob.c, and try things like "./a.out eth0", etc.)::
773
774    /* Written from scratch, but kernel-to-user space API usage
775    * dissected from lolpcap:
776    *  Copyright 2011, Chetan Loke <loke.chetan@gmail.com>
777    *  License: GPL, version 2.0
778    */
779
780    #include <stdio.h>
781    #include <stdlib.h>
782    #include <stdint.h>
783    #include <string.h>
784    #include <assert.h>
785    #include <net/if.h>
786    #include <arpa/inet.h>
787    #include <netdb.h>
788    #include <poll.h>
789    #include <unistd.h>
790    #include <signal.h>
791    #include <inttypes.h>
792    #include <sys/socket.h>
793    #include <sys/mman.h>
794    #include <linux/if_packet.h>
795    #include <linux/if_ether.h>
796    #include <linux/ip.h>
797
798    #ifndef likely
799    # define likely(x)		__builtin_expect(!!(x), 1)
800    #endif
801    #ifndef unlikely
802    # define unlikely(x)		__builtin_expect(!!(x), 0)
803    #endif
804
805    struct block_desc {
806	    uint32_t version;
807	    uint32_t offset_to_priv;
808	    struct tpacket_hdr_v1 h1;
809    };
810
811    struct ring {
812	    struct iovec *rd;
813	    uint8_t *map;
814	    struct tpacket_req3 req;
815    };
816
817    static unsigned long packets_total = 0, bytes_total = 0;
818    static sig_atomic_t sigint = 0;
819
820    static void sighandler(int num)
821    {
822	    sigint = 1;
823    }
824
825    static int setup_socket(struct ring *ring, char *netdev)
826    {
827	    int err, i, fd, v = TPACKET_V3;
828	    struct sockaddr_ll ll;
829	    unsigned int blocksiz = 1 << 22, framesiz = 1 << 11;
830	    unsigned int blocknum = 64;
831
832	    fd = socket(AF_PACKET, SOCK_RAW, htons(ETH_P_ALL));
833	    if (fd < 0) {
834		    perror("socket");
835		    exit(1);
836	    }
837
838	    err = setsockopt(fd, SOL_PACKET, PACKET_VERSION, &v, sizeof(v));
839	    if (err < 0) {
840		    perror("setsockopt");
841		    exit(1);
842	    }
843
844	    memset(&ring->req, 0, sizeof(ring->req));
845	    ring->req.tp_block_size = blocksiz;
846	    ring->req.tp_frame_size = framesiz;
847	    ring->req.tp_block_nr = blocknum;
848	    ring->req.tp_frame_nr = (blocksiz * blocknum) / framesiz;
849	    ring->req.tp_retire_blk_tov = 60;
850	    ring->req.tp_feature_req_word = TP_FT_REQ_FILL_RXHASH;
851
852	    err = setsockopt(fd, SOL_PACKET, PACKET_RX_RING, &ring->req,
853			    sizeof(ring->req));
854	    if (err < 0) {
855		    perror("setsockopt");
856		    exit(1);
857	    }
858
859	    ring->map = mmap(NULL, ring->req.tp_block_size * ring->req.tp_block_nr,
860			    PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, fd, 0);
861	    if (ring->map == MAP_FAILED) {
862		    perror("mmap");
863		    exit(1);
864	    }
865
866	    ring->rd = malloc(ring->req.tp_block_nr * sizeof(*ring->rd));
867	    assert(ring->rd);
868	    for (i = 0; i < ring->req.tp_block_nr; ++i) {
869		    ring->rd[i].iov_base = ring->map + (i * ring->req.tp_block_size);
870		    ring->rd[i].iov_len = ring->req.tp_block_size;
871	    }
872
873	    memset(&ll, 0, sizeof(ll));
874	    ll.sll_family = PF_PACKET;
875	    ll.sll_protocol = htons(ETH_P_ALL);
876	    ll.sll_ifindex = if_nametoindex(netdev);
877	    ll.sll_hatype = 0;
878	    ll.sll_pkttype = 0;
879	    ll.sll_halen = 0;
880
881	    err = bind(fd, (struct sockaddr *) &ll, sizeof(ll));
882	    if (err < 0) {
883		    perror("bind");
884		    exit(1);
885	    }
886
887	    return fd;
888    }
889
890    static void display(struct tpacket3_hdr *ppd)
891    {
892	    struct ethhdr *eth = (struct ethhdr *) ((uint8_t *) ppd + ppd->tp_mac);
893	    struct iphdr *ip = (struct iphdr *) ((uint8_t *) eth + ETH_HLEN);
894
895	    if (eth->h_proto == htons(ETH_P_IP)) {
896		    struct sockaddr_in ss, sd;
897		    char sbuff[NI_MAXHOST], dbuff[NI_MAXHOST];
898
899		    memset(&ss, 0, sizeof(ss));
900		    ss.sin_family = PF_INET;
901		    ss.sin_addr.s_addr = ip->saddr;
902		    getnameinfo((struct sockaddr *) &ss, sizeof(ss),
903				sbuff, sizeof(sbuff), NULL, 0, NI_NUMERICHOST);
904
905		    memset(&sd, 0, sizeof(sd));
906		    sd.sin_family = PF_INET;
907		    sd.sin_addr.s_addr = ip->daddr;
908		    getnameinfo((struct sockaddr *) &sd, sizeof(sd),
909				dbuff, sizeof(dbuff), NULL, 0, NI_NUMERICHOST);
910
911		    printf("%s -> %s, ", sbuff, dbuff);
912	    }
913
914	    printf("rxhash: 0x%x\n", ppd->hv1.tp_rxhash);
915    }
916
917    static void walk_block(struct block_desc *pbd, const int block_num)
918    {
919	    int num_pkts = pbd->h1.num_pkts, i;
920	    unsigned long bytes = 0;
921	    struct tpacket3_hdr *ppd;
922
923	    ppd = (struct tpacket3_hdr *) ((uint8_t *) pbd +
924					pbd->h1.offset_to_first_pkt);
925	    for (i = 0; i < num_pkts; ++i) {
926		    bytes += ppd->tp_snaplen;
927		    display(ppd);
928
929		    ppd = (struct tpacket3_hdr *) ((uint8_t *) ppd +
930						ppd->tp_next_offset);
931	    }
932
933	    packets_total += num_pkts;
934	    bytes_total += bytes;
935    }
936
937    static void flush_block(struct block_desc *pbd)
938    {
939	    pbd->h1.block_status = TP_STATUS_KERNEL;
940    }
941
942    static void teardown_socket(struct ring *ring, int fd)
943    {
944	    munmap(ring->map, ring->req.tp_block_size * ring->req.tp_block_nr);
945	    free(ring->rd);
946	    close(fd);
947    }
948
949    int main(int argc, char **argp)
950    {
951	    int fd, err;
952	    socklen_t len;
953	    struct ring ring;
954	    struct pollfd pfd;
955	    unsigned int block_num = 0, blocks = 64;
956	    struct block_desc *pbd;
957	    struct tpacket_stats_v3 stats;
958
959	    if (argc != 2) {
960		    fprintf(stderr, "Usage: %s INTERFACE\n", argp[0]);
961		    return EXIT_FAILURE;
962	    }
963
964	    signal(SIGINT, sighandler);
965
966	    memset(&ring, 0, sizeof(ring));
967	    fd = setup_socket(&ring, argp[argc - 1]);
968	    assert(fd > 0);
969
970	    memset(&pfd, 0, sizeof(pfd));
971	    pfd.fd = fd;
972	    pfd.events = POLLIN | POLLERR;
973	    pfd.revents = 0;
974
975	    while (likely(!sigint)) {
976		    pbd = (struct block_desc *) ring.rd[block_num].iov_base;
977
978		    if ((pbd->h1.block_status & TP_STATUS_USER) == 0) {
979			    poll(&pfd, 1, -1);
980			    continue;
981		    }
982
983		    walk_block(pbd, block_num);
984		    flush_block(pbd);
985		    block_num = (block_num + 1) % blocks;
986	    }
987
988	    len = sizeof(stats);
989	    err = getsockopt(fd, SOL_PACKET, PACKET_STATISTICS, &stats, &len);
990	    if (err < 0) {
991		    perror("getsockopt");
992		    exit(1);
993	    }
994
995	    fflush(stdout);
996	    printf("\nReceived %u packets, %lu bytes, %u dropped, freeze_q_cnt: %u\n",
997		stats.tp_packets, bytes_total, stats.tp_drops,
998		stats.tp_freeze_q_cnt);
999
1000	    teardown_socket(&ring, fd);
1001	    return 0;
1002    }
1003
1004PACKET_QDISC_BYPASS
1005===================
1006
1007If there is a requirement to load the network with many packets in a similar
1008fashion as pktgen does, you might set the following option after socket
1009creation::
1010
1011    int one = 1;
1012    setsockopt(fd, SOL_PACKET, PACKET_QDISC_BYPASS, &one, sizeof(one));
1013
1014This has the side-effect, that packets sent through PF_PACKET will bypass the
1015kernel's qdisc layer and are forcedly pushed to the driver directly. Meaning,
1016packet are not buffered, tc disciplines are ignored, increased loss can occur
1017and such packets are also not visible to other PF_PACKET sockets anymore. So,
1018you have been warned; generally, this can be useful for stress testing various
1019components of a system.
1020
1021On default, PACKET_QDISC_BYPASS is disabled and needs to be explicitly enabled
1022on PF_PACKET sockets.
1023
1024PACKET_TIMESTAMP
1025================
1026
1027The PACKET_TIMESTAMP setting determines the source of the timestamp in
1028the packet meta information for mmap(2)ed RX_RING and TX_RINGs.  If your
1029NIC is capable of timestamping packets in hardware, you can request those
1030hardware timestamps to be used. Note: you may need to enable the generation
1031of hardware timestamps with SIOCSHWTSTAMP (see related information from
1032Documentation/networking/timestamping.rst).
1033
1034PACKET_TIMESTAMP accepts the same integer bit field as SO_TIMESTAMPING::
1035
1036    int req = SOF_TIMESTAMPING_RAW_HARDWARE;
1037    setsockopt(fd, SOL_PACKET, PACKET_TIMESTAMP, (void *) &req, sizeof(req))
1038
1039For the mmap(2)ed ring buffers, such timestamps are stored in the
1040``tpacket{,2,3}_hdr`` structure's tp_sec and ``tp_{n,u}sec`` members.
1041To determine what kind of timestamp has been reported, the tp_status field
1042is binary or'ed with the following possible bits ...
1043
1044::
1045
1046    TP_STATUS_TS_RAW_HARDWARE
1047    TP_STATUS_TS_SOFTWARE
1048
1049... that are equivalent to its ``SOF_TIMESTAMPING_*`` counterparts. For the
1050RX_RING, if neither is set (i.e. PACKET_TIMESTAMP is not set), then a
1051software fallback was invoked *within* PF_PACKET's processing code (less
1052precise).
1053
1054Getting timestamps for the TX_RING works as follows: i) fill the ring frames,
1055ii) call sendto() e.g. in blocking mode, iii) wait for status of relevant
1056frames to be updated resp. the frame handed over to the application, iv) walk
1057through the frames to pick up the individual hw/sw timestamps.
1058
1059Only (!) if transmit timestamping is enabled, then these bits are combined
1060with binary | with TP_STATUS_AVAILABLE, so you must check for that in your
1061application (e.g. !(tp_status & (TP_STATUS_SEND_REQUEST | TP_STATUS_SENDING))
1062in a first step to see if the frame belongs to the application, and then
1063one can extract the type of timestamp in a second step from tp_status)!
1064
1065If you don't care about them, thus having it disabled, checking for
1066TP_STATUS_AVAILABLE resp. TP_STATUS_WRONG_FORMAT is sufficient. If in the
1067TX_RING part only TP_STATUS_AVAILABLE is set, then the tp_sec and tp_{n,u}sec
1068members do not contain a valid value. For TX_RINGs, by default no timestamp
1069is generated!
1070
1071See include/linux/net_tstamp.h and Documentation/networking/timestamping.rst
1072for more information on hardware timestamps.
1073
1074Miscellaneous bits
1075==================
1076
1077- Packet sockets work well together with Linux socket filters, thus you also
1078  might want to have a look at Documentation/networking/filter.rst
1079
1080THANKS
1081======
1082
1083   Jesse Brandeburg, for fixing my grammathical/spelling errors
1084