linux/Documentation/networking/packet_mmap.txt
<<
>>
Prefs
   1--------------------------------------------------------------------------------
   2+ ABSTRACT
   3--------------------------------------------------------------------------------
   4
   5This file documents the mmap() facility available with the PACKET
   6socket interface on 2.4 and 2.6 kernels. This type of sockets is used for 
   7capture network traffic with utilities like tcpdump or any other that needs
   8raw access to network interface.
   9
  10You can find the latest version of this document at:
  11    http://wiki.ipxwarzone.com/index.php5?title=Linux_packet_mmap
  12
  13Howto can be found at:
  14    http://wiki.gnu-log.net (packet_mmap)
  15
  16Please send your comments to
  17    Ulisses Alonso CamarĂ³ <uaca@i.hate.spam.alumni.uv.es>
  18    Johann Baudy <johann.baudy@gnu-log.net>
  19
  20-------------------------------------------------------------------------------
  21+ Why use PACKET_MMAP
  22--------------------------------------------------------------------------------
  23
  24In Linux 2.4/2.6 if PACKET_MMAP is not enabled, the capture process is very
  25inefficient. It uses very limited buffers and requires one system call
  26to capture each packet, it requires two if you want to get packet's 
  27timestamp (like libpcap always does).
  28
  29In the other hand PACKET_MMAP is very efficient. PACKET_MMAP provides a size 
  30configurable circular buffer mapped in user space that can be used to either
  31send or receive packets. This way reading packets just needs to wait for them,
  32most of the time there is no need to issue a single system call. Concerning
  33transmission, multiple packets can be sent through one system call to get the
  34highest bandwidth.
  35By using a shared buffer between the kernel and the user also has the benefit
  36of minimizing packet copies.
  37
  38It's fine to use PACKET_MMAP to improve the performance of the capture and
  39transmission process, but it isn't everything. At least, if you are capturing
  40at high speeds (this is relative to the cpu speed), you should check if the
  41device driver of your network interface card supports some sort of interrupt
  42load mitigation or (even better) if it supports NAPI, also make sure it is
  43enabled. For transmission, check the MTU (Maximum Transmission Unit) used and
  44supported by devices of your network.
  45
  46--------------------------------------------------------------------------------
  47+ How to use mmap() to improve capture process
  48--------------------------------------------------------------------------------
  49
  50From the user standpoint, you should use the higher level libpcap library, which
  51is a de facto standard, portable across nearly all operating systems
  52including Win32. 
  53
  54Said that, at time of this writing, official libpcap 0.8.1 is out and doesn't include
  55support for PACKET_MMAP, and also probably the libpcap included in your distribution. 
  56
  57I'm aware of two implementations of PACKET_MMAP in libpcap:
  58
  59    http://wiki.ipxwarzone.com/              (by Simon Patarin, based on libpcap 0.6.2)
  60    http://public.lanl.gov/cpw/              (by Phil Wood, based on lastest libpcap)
  61
  62The rest of this document is intended for people who want to understand
  63the low level details or want to improve libpcap by including PACKET_MMAP
  64support.
  65
  66--------------------------------------------------------------------------------
  67+ How to use mmap() directly to improve capture process
  68--------------------------------------------------------------------------------
  69
  70From the system calls stand point, the use of PACKET_MMAP involves
  71the following process:
  72
  73
  74[setup]     socket() -------> creation of the capture socket
  75            setsockopt() ---> allocation of the circular buffer (ring)
  76                              option: PACKET_RX_RING
  77            mmap() ---------> mapping of the allocated buffer to the
  78                              user process
  79
  80[capture]   poll() ---------> to wait for incoming packets
  81
  82[shutdown]  close() --------> destruction of the capture socket and
  83                              deallocation of all associated 
  84                              resources.
  85
  86
  87socket creation and destruction is straight forward, and is done 
  88the same way with or without PACKET_MMAP:
  89
  90int fd;
  91
  92fd= socket(PF_PACKET, mode, htons(ETH_P_ALL))
  93
  94where mode is SOCK_RAW for the raw interface were link level
  95information can be captured or SOCK_DGRAM for the cooked
  96interface where link level information capture is not 
  97supported and a link level pseudo-header is provided 
  98by the kernel.
  99
 100The destruction of the socket and all associated resources
 101is done by a simple call to close(fd).
 102
 103Next I will describe PACKET_MMAP settings and its constraints,
 104also the mapping of the circular buffer in the user process and 
 105the use of this buffer.
 106
 107--------------------------------------------------------------------------------
 108+ How to use mmap() directly to improve transmission process
 109--------------------------------------------------------------------------------
 110Transmission process is similar to capture as shown below.
 111
 112[setup]          socket() -------> creation of the transmission socket
 113                 setsockopt() ---> allocation of the circular buffer (ring)
 114                                   option: PACKET_TX_RING
 115                 bind() ---------> bind transmission socket with a network interface
 116                 mmap() ---------> mapping of the allocated buffer to the
 117                                   user process
 118
 119[transmission]   poll() ---------> wait for free packets (optional)
 120                 send() ---------> send all packets that are set as ready in
 121                                   the ring
 122                                   The flag MSG_DONTWAIT can be used to return
 123                                   before end of transfer.
 124
 125[shutdown]  close() --------> destruction of the transmission socket and
 126                              deallocation of all associated resources.
 127
 128Binding the socket to your network interface is mandatory (with zero copy) to
 129know the header size of frames used in the circular buffer.
 130
 131As capture, each frame contains two parts:
 132
 133 --------------------
 134| struct tpacket_hdr | Header. It contains the status of
 135|                    | of this frame
 136|--------------------|
 137| data buffer        |
 138.                    .  Data that will be sent over the network interface.
 139.                    .
 140 --------------------
 141
 142 bind() associates the socket to your network interface thanks to
 143 sll_ifindex parameter of struct sockaddr_ll.
 144
 145 Initialization example:
 146
 147 struct sockaddr_ll my_addr;
 148 struct ifreq s_ifr;
 149 ...
 150
 151 strncpy (s_ifr.ifr_name, "eth0", sizeof(s_ifr.ifr_name));
 152
 153 /* get interface index of eth0 */
 154 ioctl(this->socket, SIOCGIFINDEX, &s_ifr);
 155
 156 /* fill sockaddr_ll struct to prepare binding */
 157 my_addr.sll_family = AF_PACKET;
 158 my_addr.sll_protocol = htons(ETH_P_ALL);
 159 my_addr.sll_ifindex =  s_ifr.ifr_ifindex;
 160
 161 /* bind socket to eth0 */
 162 bind(this->socket, (struct sockaddr *)&my_addr, sizeof(struct sockaddr_ll));
 163
 164 A complete tutorial is available at: http://wiki.gnu-log.net/
 165
 166--------------------------------------------------------------------------------
 167+ PACKET_MMAP settings
 168--------------------------------------------------------------------------------
 169
 170
 171To setup PACKET_MMAP from user level code is done with a call like
 172
 173 - Capture process
 174     setsockopt(fd, SOL_PACKET, PACKET_RX_RING, (void *) &req, sizeof(req))
 175 - Transmission process
 176     setsockopt(fd, SOL_PACKET, PACKET_TX_RING, (void *) &req, sizeof(req))
 177
 178The most significant argument in the previous call is the req parameter, 
 179this parameter must to have the following structure:
 180
 181    struct tpacket_req
 182    {
 183        unsigned int    tp_block_size;  /* Minimal size of contiguous block */
 184        unsigned int    tp_block_nr;    /* Number of blocks */
 185        unsigned int    tp_frame_size;  /* Size of frame */
 186        unsigned int    tp_frame_nr;    /* Total number of frames */
 187    };
 188
 189This structure is defined in /usr/include/linux/if_packet.h and establishes a 
 190circular buffer (ring) of unswappable memory.
 191Being mapped in the capture process allows reading the captured frames and 
 192related meta-information like timestamps without requiring a system call.
 193
 194Frames are grouped in blocks. Each block is a physically contiguous
 195region of memory and holds tp_block_size/tp_frame_size frames. The total number 
 196of blocks is tp_block_nr. Note that tp_frame_nr is a redundant parameter because
 197
 198    frames_per_block = tp_block_size/tp_frame_size
 199
 200indeed, packet_set_ring checks that the following condition is true
 201
 202    frames_per_block * tp_block_nr == tp_frame_nr
 203
 204
 205Lets see an example, with the following values:
 206
 207     tp_block_size= 4096
 208     tp_frame_size= 2048
 209     tp_block_nr  = 4
 210     tp_frame_nr  = 8
 211
 212we will get the following buffer structure:
 213
 214        block #1                 block #2         
 215+---------+---------+    +---------+---------+    
 216| frame 1 | frame 2 |    | frame 3 | frame 4 |    
 217+---------+---------+    +---------+---------+    
 218
 219        block #3                 block #4
 220+---------+---------+    +---------+---------+
 221| frame 5 | frame 6 |    | frame 7 | frame 8 |
 222+---------+---------+    +---------+---------+
 223
 224A frame can be of any size with the only condition it can fit in a block. A block
 225can only hold an integer number of frames, or in other words, a frame cannot 
 226be spawned across two blocks, so there are some details you have to take into 
 227account when choosing the frame_size. See "Mapping and use of the circular 
 228buffer (ring)".
 229
 230
 231--------------------------------------------------------------------------------
 232+ PACKET_MMAP setting constraints
 233--------------------------------------------------------------------------------
 234
 235In kernel versions prior to 2.4.26 (for the 2.4 branch) and 2.6.5 (2.6 branch),
 236the PACKET_MMAP buffer could hold only 32768 frames in a 32 bit architecture or
 23716384 in a 64 bit architecture. For information on these kernel versions
 238see http://pusa.uv.es/~ulisses/packet_mmap/packet_mmap.pre-2.4.26_2.6.5.txt
 239
 240 Block size limit
 241------------------
 242
 243As stated earlier, each block is a contiguous physical region of memory. These 
 244memory regions are allocated with calls to the __get_free_pages() function. As 
 245the name indicates, this function allocates pages of memory, and the second
 246argument is "order" or a power of two number of pages, that is 
 247(for PAGE_SIZE == 4096) order=0 ==> 4096 bytes, order=1 ==> 8192 bytes, 
 248order=2 ==> 16384 bytes, etc. The maximum size of a 
 249region allocated by __get_free_pages is determined by the MAX_ORDER macro. More 
 250precisely the limit can be calculated as:
 251
 252   PAGE_SIZE << MAX_ORDER
 253
 254   In a i386 architecture PAGE_SIZE is 4096 bytes 
 255   In a 2.4/i386 kernel MAX_ORDER is 10
 256   In a 2.6/i386 kernel MAX_ORDER is 11
 257
 258So get_free_pages can allocate as much as 4MB or 8MB in a 2.4/2.6 kernel 
 259respectively, with an i386 architecture.
 260
 261User space programs can include /usr/include/sys/user.h and 
 262/usr/include/linux/mmzone.h to get PAGE_SIZE MAX_ORDER declarations.
 263
 264The pagesize can also be determined dynamically with the getpagesize (2) 
 265system call. 
 266
 267
 268 Block number limit
 269--------------------
 270
 271To understand the constraints of PACKET_MMAP, we have to see the structure 
 272used to hold the pointers to each block.
 273
 274Currently, this structure is a dynamically allocated vector with kmalloc 
 275called pg_vec, its size limits the number of blocks that can be allocated.
 276
 277    +---+---+---+---+
 278    | x | x | x | x |
 279    +---+---+---+---+
 280      |   |   |   |
 281      |   |   |   v
 282      |   |   v  block #4
 283      |   v  block #3
 284      v  block #2
 285     block #1
 286
 287
 288kmalloc allocates any number of bytes of physically contiguous memory from 
 289a pool of pre-determined sizes. This pool of memory is maintained by the slab 
 290allocator which is at the end the responsible for doing the allocation and 
 291hence which imposes the maximum memory that kmalloc can allocate. 
 292
 293In a 2.4/2.6 kernel and the i386 architecture, the limit is 131072 bytes. The 
 294predetermined sizes that kmalloc uses can be checked in the "size-<bytes>" 
 295entries of /proc/slabinfo
 296
 297In a 32 bit architecture, pointers are 4 bytes long, so the total number of 
 298pointers to blocks is
 299
 300     131072/4 = 32768 blocks
 301
 302
 303 PACKET_MMAP buffer size calculator
 304------------------------------------
 305
 306Definitions:
 307
 308<size-max>    : is the maximum size of allocable with kmalloc (see /proc/slabinfo)
 309<pointer size>: depends on the architecture -- sizeof(void *)
 310<page size>   : depends on the architecture -- PAGE_SIZE or getpagesize (2)
 311<max-order>   : is the value defined with MAX_ORDER
 312<frame size>  : it's an upper bound of frame's capture size (more on this later)
 313
 314from these definitions we will derive 
 315
 316        <block number> = <size-max>/<pointer size>
 317        <block size> = <pagesize> << <max-order>
 318
 319so, the max buffer size is
 320
 321        <block number> * <block size>
 322
 323and, the number of frames be
 324
 325        <block number> * <block size> / <frame size>
 326
 327Suppose the following parameters, which apply for 2.6 kernel and an
 328i386 architecture:
 329
 330        <size-max> = 131072 bytes
 331        <pointer size> = 4 bytes
 332        <pagesize> = 4096 bytes
 333        <max-order> = 11
 334
 335and a value for <frame size> of 2048 bytes. These parameters will yield
 336
 337        <block number> = 131072/4 = 32768 blocks
 338        <block size> = 4096 << 11 = 8 MiB.
 339
 340and hence the buffer will have a 262144 MiB size. So it can hold 
 341262144 MiB / 2048 bytes = 134217728 frames
 342
 343
 344Actually, this buffer size is not possible with an i386 architecture. 
 345Remember that the memory is allocated in kernel space, in the case of 
 346an i386 kernel's memory size is limited to 1GiB.
 347
 348All memory allocations are not freed until the socket is closed. The memory 
 349allocations are done with GFP_KERNEL priority, this basically means that 
 350the allocation can wait and swap other process' memory in order to allocate 
 351the necessary memory, so normally limits can be reached.
 352
 353 Other constraints
 354-------------------
 355
 356If you check the source code you will see that what I draw here as a frame
 357is not only the link level frame. At the beginning of each frame there is a 
 358header called struct tpacket_hdr used in PACKET_MMAP to hold link level's frame
 359meta information like timestamp. So what we draw here a frame it's really 
 360the following (from include/linux/if_packet.h):
 361
 362/*
 363   Frame structure:
 364
 365   - Start. Frame must be aligned to TPACKET_ALIGNMENT=16
 366   - struct tpacket_hdr
 367   - pad to TPACKET_ALIGNMENT=16
 368   - struct sockaddr_ll
 369   - Gap, chosen so that packet data (Start+tp_net) aligns to 
 370     TPACKET_ALIGNMENT=16
 371   - Start+tp_mac: [ Optional MAC header ]
 372   - Start+tp_net: Packet data, aligned to TPACKET_ALIGNMENT=16.
 373   - Pad to align to TPACKET_ALIGNMENT=16
 374 */
 375           
 376 
 377 The following are conditions that are checked in packet_set_ring
 378
 379   tp_block_size must be a multiple of PAGE_SIZE (1)
 380   tp_frame_size must be greater than TPACKET_HDRLEN (obvious)
 381   tp_frame_size must be a multiple of TPACKET_ALIGNMENT
 382   tp_frame_nr   must be exactly frames_per_block*tp_block_nr
 383
 384Note that tp_block_size should be chosen to be a power of two or there will
 385be a waste of memory.
 386
 387--------------------------------------------------------------------------------
 388+ Mapping and use of the circular buffer (ring)
 389--------------------------------------------------------------------------------
 390
 391The mapping of the buffer in the user process is done with the conventional 
 392mmap function. Even the circular buffer is compound of several physically
 393discontiguous blocks of memory, they are contiguous to the user space, hence
 394just one call to mmap is needed:
 395
 396    mmap(0, size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
 397
 398If tp_frame_size is a divisor of tp_block_size frames will be 
 399contiguously spaced by tp_frame_size bytes. If not, each
 400tp_block_size/tp_frame_size frames there will be a gap between 
 401the frames. This is because a frame cannot be spawn across two
 402blocks. 
 403
 404At the beginning of each frame there is an status field (see 
 405struct tpacket_hdr). If this field is 0 means that the frame is ready
 406to be used for the kernel, If not, there is a frame the user can read 
 407and the following flags apply:
 408
 409+++ Capture process:
 410     from include/linux/if_packet.h
 411
 412     #define TP_STATUS_COPY          2 
 413     #define TP_STATUS_LOSING        4 
 414     #define TP_STATUS_CSUMNOTREADY  8 
 415
 416
 417TP_STATUS_COPY        : This flag indicates that the frame (and associated
 418                        meta information) has been truncated because it's 
 419                        larger than tp_frame_size. This packet can be 
 420                        read entirely with recvfrom().
 421                        
 422                        In order to make this work it must to be
 423                        enabled previously with setsockopt() and 
 424                        the PACKET_COPY_THRESH option. 
 425
 426                        The number of frames than can be buffered to 
 427                        be read with recvfrom is limited like a normal socket.
 428                        See the SO_RCVBUF option in the socket (7) man page.
 429
 430TP_STATUS_LOSING      : indicates there were packet drops from last time 
 431                        statistics where checked with getsockopt() and
 432                        the PACKET_STATISTICS option.
 433
 434TP_STATUS_CSUMNOTREADY: currently it's used for outgoing IP packets which 
 435                        its checksum will be done in hardware. So while
 436                        reading the packet we should not try to check the 
 437                        checksum. 
 438
 439for convenience there are also the following defines:
 440
 441     #define TP_STATUS_KERNEL        0
 442     #define TP_STATUS_USER          1
 443
 444The kernel initializes all frames to TP_STATUS_KERNEL, when the kernel
 445receives a packet it puts in the buffer and updates the status with
 446at least the TP_STATUS_USER flag. Then the user can read the packet,
 447once the packet is read the user must zero the status field, so the kernel 
 448can use again that frame buffer.
 449
 450The user can use poll (any other variant should apply too) to check if new
 451packets are in the ring:
 452
 453    struct pollfd pfd;
 454
 455    pfd.fd = fd;
 456    pfd.revents = 0;
 457    pfd.events = POLLIN|POLLRDNORM|POLLERR;
 458
 459    if (status == TP_STATUS_KERNEL)
 460        retval = poll(&pfd, 1, timeout);
 461
 462It doesn't incur in a race condition to first check the status value and 
 463then poll for frames.
 464
 465
 466++ Transmission process
 467Those defines are also used for transmission:
 468
 469     #define TP_STATUS_AVAILABLE        0 // Frame is available
 470     #define TP_STATUS_SEND_REQUEST     1 // Frame will be sent on next send()
 471     #define TP_STATUS_SENDING          2 // Frame is currently in transmission
 472     #define TP_STATUS_WRONG_FORMAT     4 // Frame format is not correct
 473
 474First, the kernel initializes all frames to TP_STATUS_AVAILABLE. To send a
 475packet, the user fills a data buffer of an available frame, sets tp_len to
 476current data buffer size and sets its status field to TP_STATUS_SEND_REQUEST.
 477This can be done on multiple frames. Once the user is ready to transmit, it
 478calls send(). Then all buffers with status equal to TP_STATUS_SEND_REQUEST are
 479forwarded to the network device. The kernel updates each status of sent
 480frames with TP_STATUS_SENDING until the end of transfer.
 481At the end of each transfer, buffer status returns to TP_STATUS_AVAILABLE.
 482
 483    header->tp_len = in_i_size;
 484    header->tp_status = TP_STATUS_SEND_REQUEST;
 485    retval = send(this->socket, NULL, 0, 0);
 486
 487The user can also use poll() to check if a buffer is available:
 488(status == TP_STATUS_SENDING)
 489
 490    struct pollfd pfd;
 491    pfd.fd = fd;
 492    pfd.revents = 0;
 493    pfd.events = POLLOUT;
 494    retval = poll(&pfd, 1, timeout);
 495
 496-------------------------------------------------------------------------------
 497+ PACKET_TIMESTAMP
 498-------------------------------------------------------------------------------
 499
 500The PACKET_TIMESTAMP setting determines the source of the timestamp in
 501the packet meta information.  If your NIC is capable of timestamping
 502packets in hardware, you can request those hardware timestamps to used.
 503Note: you may need to enable the generation of hardware timestamps with
 504SIOCSHWTSTAMP.
 505
 506PACKET_TIMESTAMP accepts the same integer bit field as
 507SO_TIMESTAMPING.  However, only the SOF_TIMESTAMPING_SYS_HARDWARE
 508and SOF_TIMESTAMPING_RAW_HARDWARE values are recognized by
 509PACKET_TIMESTAMP.  SOF_TIMESTAMPING_SYS_HARDWARE takes precedence over
 510SOF_TIMESTAMPING_RAW_HARDWARE if both bits are set.
 511
 512    int req = 0;
 513    req |= SOF_TIMESTAMPING_SYS_HARDWARE;
 514    setsockopt(fd, SOL_PACKET, PACKET_TIMESTAMP, (void *) &req, sizeof(req))
 515
 516If PACKET_TIMESTAMP is not set, a software timestamp generated inside
 517the networking stack is used (the behavior before this setting was added).
 518
 519See include/linux/net_tstamp.h and Documentation/networking/timestamping
 520for more information on hardware timestamps.
 521
 522--------------------------------------------------------------------------------
 523+ THANKS
 524--------------------------------------------------------------------------------
 525   
 526   Jesse Brandeburg, for fixing my grammathical/spelling errors
 527
 528
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.