--------------------------------------------------------------------------------
+ ABSTRACT
--------------------------------------------------------------------------------

This file documents the CONFIG_PACKET_MMAP option available with the PACKET
socket interface on 2.4 and 2.6 kernels. This type of sockets is used for 
capture network traffic with utilities like tcpdump or any other that uses 
the libpcap library. 

You can find the latest version of this document at

    http://pusa.uv.es/~ulisses/packet_mmap/

Please send me your comments to

    Ulisses Alonso Camaró <uaca@i.hate.spam.alumni.uv.es>

-------------------------------------------------------------------------------
+ Why use PACKET_MMAP
--------------------------------------------------------------------------------

In Linux 2.4/2.6 if PACKET_MMAP is not enabled, the capture process is very
inefficient. It uses very limited buffers and requires one system call
to capture each packet, it requires two if you want to get packet's 
timestamp (like libpcap always does).

In the other hand PACKET_MMAP is very efficient. PACKET_MMAP provides a size 
configurable circular buffer mapped in user space. This way reading packets just 
needs to wait for them, most of the time there is no need to issue a single 
system call. By using a shared buffer between the kernel and the user 
also has the benefit of minimizing packet copies.

It's fine to use PACKET_MMAP to improve the performance of the capture process, 
but it isn't everything. At least, if you are capturing at high speeds (this 
is relative to the cpu speed), you should check if the device driver of your 
network interface card supports some sort of interrupt load mitigation or 
(even better) if it supports NAPI, also make sure it is enabled.

--------------------------------------------------------------------------------
+ How to use CONFIG_PACKET_MMAP
--------------------------------------------------------------------------------

From the user standpoint, you should use the higher level libpcap library, which
is a de facto standard, portable across nearly all operating systems
including Win32. 

Said that, at time of this writing, official libpcap 0.8.1 is out and doesn't include
support for PACKET_MMAP, and also probably the libpcap included in your distribution. 

I'm aware of two implementations of PACKET_MMAP in libpcap:

    http://pusa.uv.es/~ulisses/packet_mmap/  (by Simon Patarin, based on libpcap 0.6.2)
    http://public.lanl.gov/cpw/              (by Phil Wood, based on lastest libpcap)

The rest of this document is intended for people who want to understand
the low level details or want to improve libpcap by including PACKET_MMAP
support.

--------------------------------------------------------------------------------
+ How to use CONFIG_PACKET_MMAP directly
--------------------------------------------------------------------------------

From the system calls stand point, the use of PACKET_MMAP involves
the following process:


[setup]     socket() -------> creation of the capture socket
            setsockopt() ---> allocation of the circular buffer (ring)
            mmap() ---------> mapping of the allocated buffer to the
                              user process

[capture]   poll() ---------> to wait for incoming packets

[shutdown]  close() --------> destruction of the capture socket and
                              deallocation of all associated 
                              resources.


socket creation and destruction is straight forward, and is done 
the same way with or without PACKET_MMAP:

int fd;

fd= socket(PF_PACKET, mode, htons(ETH_P_ALL))

where mode is SOCK_RAW for the raw interface were link level
information can be captured or SOCK_DGRAM for the cooked
interface where link level information capture is not 
supported and a link level pseudo-header is provided 
by the kernel.

The destruction of the socket and all associated resources
is done by a simple call to close(fd).

Next I will describe PACKET_MMAP settings and it's constraints,
also the mapping of the circular buffer in the user process and 
the use of this buffer.

--------------------------------------------------------------------------------
+ PACKET_MMAP settings
--------------------------------------------------------------------------------


To setup PACKET_MMAP from user level code is done with a call like

     setsockopt(fd, SOL_PACKET, PACKET_RX_RING, (void *) &req, sizeof(req))

The most significant argument in the previous call is the req parameter, 
this parameter must to have the following structure:

    struct tpacket_req
    {
        unsigned int    tp_block_size;  /* Minimal size of contiguous block */
        unsigned int    tp_block_nr;    /* Number of blocks */
        unsigned int    tp_frame_size;  /* Size of frame */
        unsigned int    tp_frame_nr;    /* Total number of frames */
    };

This structure is defined in /usr/include/linux/if_packet.h and establishes a 
circular buffer (ring) of unswappable memory mapped in the capture process. 
Being mapped in the capture process allows reading the captured frames and 
related meta-information like timestamps without requiring a system call.

Captured frames are grouped in blocks. Each block is a physically contiguous 
region of memory and holds tp_block_size/tp_frame_size frames. The total number 
of blocks is tp_block_nr. Note that tp_frame_nr is a redundant parameter because

    frames_per_block = tp_block_size/tp_frame_size

indeed, packet_set_ring checks that the following condition is true

    frames_per_block * tp_block_nr == tp_frame_nr


Lets see an example, with the following values:

     tp_block_size= 4096
     tp_frame_size= 2048
     tp_block_nr  = 4
     tp_frame_nr  = 8

we will get the following buffer structure:

        block #1                 block #2         
+---------+---------+    +---------+---------+    
| frame 1 | frame 2 |    | frame 3 | frame 4 |    
+---------+---------+    +---------+---------+    

        block #3                 block #4
+---------+---------+    +---------+---------+
| frame 5 | frame 6 |    | frame 7 | frame 8 |
+---------+---------+    +---------+---------+

A frame can be of any size with the only condition it can fit in a block. A block
can only hold an integer number of frames, or in other words, a frame cannot 
be spawned accross two blocks, so there are some details you have to take into 
account when choosing the frame_size. See "Mapping and use of the circular 
buffer (ring)".


--------------------------------------------------------------------------------
+ PACKET_MMAP setting constraints
--------------------------------------------------------------------------------

In kernel versions prior to 2.4.26 (for the 2.4 branch) and 2.6.5 (2.6 branch),
the PACKET_MMAP buffer could hold only 32768 frames in a 32 bit architecture or
16384 in a 64 bit architecture. For information on these kernel versions
see http://pusa.uv.es/~ulisses/packet_mmap/packet_mmap.pre-2.4.26_2.6.5.txt

 Block size limit
------------------

As stated earlier, each block is a contiguous physical region of memory. These 
memory regions are allocated with calls to the __get_free_pages() function. As 
the name indicates, this function allocates pages of memory, and the second
argument is "order" or a power of two number of pages, that is 
(for PAGE_SIZE == 4096) order=0 ==> 4096 bytes, order=1 ==> 8192 bytes, 
order=2 ==> 16384 bytes, etc. The maximum size of a 
region allocated by __get_free_pages is determined by the MAX_ORDER macro. More 
precisely the limit can be calculated as:

   PAGE_SIZE << MAX_ORDER

   In a i386 architecture PAGE_SIZE is 4096 bytes 
   In a 2.4/i386 kernel MAX_ORDER is 10
   In a 2.6/i386 kernel MAX_ORDER is 11

So get_free_pages can allocate as much as 4MB or 8MB in a 2.4/2.6 kernel 
respectively, with an i386 architecture.

User space programs can include /usr/include/sys/user.h and 
/usr/include/linux/mmzone.h to get PAGE_SIZE MAX_ORDER declarations.

The pagesize can also be determined dynamically with the getpagesize (2) 
system call. 


 Block number limit
--------------------

To understand the constraints of PACKET_MMAP, we have to see the structure 
used to hold the pointers to each block.

Currently, this structure is a dynamically allocated vector with kmalloc 
called pg_vec, its size limits the number of blocks that can be allocated.

    +---+---+---+---+
    | x | x | x | x |
    +---+---+---+---+
      |   |   |   |
      |   |   |   v
      |   |   v  block #4
      |   v  block #3
      v  block #2
     block #1


kmalloc allocates any number of bytes of physically contiguous memory from 
a pool of pre-determined sizes. This pool of memory is maintained by the slab 
allocator which is at the end the responsible for doing the allocation and 
hence which imposes the maximum memory that kmalloc can allocate. 

In a 2.4/2.6 kernel and the i386 architecture, the limit is 131072 bytes. The 
predetermined sizes that kmalloc uses can be checked in the "size-<bytes>" 
entries of /proc/slabinfo

In a 32 bit architecture, pointers are 4 bytes long, so the total number of 
pointers to blocks is

     131072/4 = 32768 blocks


 PACKET_MMAP buffer size calculator
------------------------------------

Definitions:

<size-max>    : is the maximum size of allocable with kmalloc (see /proc/slabinfo)
<pointer size>: depends on the architecture -- sizeof(void *)
<page size>   : depends on the architecture -- PAGE_SIZE or getpagesize (2)
<max-order>   : is the value defined with MAX_ORDER
<frame size>  : it's an upper bound of frame's capture size (more on this later)

from these definitions we will derive 

	<block number> = <size-max>/<pointer size>
	<block size> = <pagesize> << <max-order>

so, the max buffer size is

	<block number> * <block size>

and, the number of frames be

	<block number> * <block size> / <frame size>

Suppose the following parameters, which apply for 2.6 kernel and an
i386 architecture:

	<size-max> = 131072 bytes
	<pointer size> = 4 bytes
	<pagesize> = 4096 bytes
	<max-order> = 11

and a value for <frame size> of 2048 bytes. These parameters will yield

	<block number> = 131072/4 = 32768 blocks
	<block size> = 4096 << 11 = 8 MiB.

and hence the buffer will have a 262144 MiB size. So it can hold 
262144 MiB / 2048 bytes = 134217728 frames


Actually, this buffer size is not possible with an i386 architecture. 
Remember that the memory is allocated in kernel space, in the case of 
an i386 kernel's memory size is limited to 1GiB.

All memory allocations are not freed until the socket is closed. The memory 
allocations are done with GFP_KERNEL priority, this basically means that 
the allocation can wait and swap other process' memory in order to allocate 
the necessary memory, so normally limits can be reached.

 Other constraints
-------------------

If you check the source code you will see that what I draw here as a frame
is not only the link level frame. At the beginning of each frame there is a 
header called struct tpacket_hdr used in PACKET_MMAP to hold link level's frame
meta information like timestamp. So what we draw here a frame it's really 
the following (from include/linux/if_packet.h):

/*
   Frame structure:

   - Start. Frame must be aligned to TPACKET_ALIGNMENT=16
   - struct tpacket_hdr
   - pad to TPACKET_ALIGNMENT=16
   - struct sockaddr_ll
   - Gap, chosen so that packet data (Start+tp_net) aligns to 
     TPACKET_ALIGNMENT=16
   - Start+tp_mac: [ Optional MAC header ]
   - Start+tp_net: Packet data, aligned to TPACKET_ALIGNMENT=16.
   - Pad to align to TPACKET_ALIGNMENT=16
 */
           
 
 The following are conditions that are checked in packet_set_ring

   tp_block_size must be a multiple of PAGE_SIZE (1)
   tp_frame_size must be greater than TPACKET_HDRLEN (obvious)
   tp_frame_size must be a multiple of TPACKET_ALIGNMENT
   tp_frame_nr   must be exactly frames_per_block*tp_block_nr

Note that tp_block_size should be chosen to be a power of two or there will
be a waste of memory.

--------------------------------------------------------------------------------
+ Mapping and use of the circular buffer (ring)
--------------------------------------------------------------------------------

The mapping of the buffer in the user process is done with the conventional 
mmap function. Even the circular buffer is compound of several physically
discontiguous blocks of memory, they are contiguous to the user space, hence
just one call to mmap is needed:

    mmap(0, size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);

If tp_frame_size is a divisor of tp_block_size frames will be 
contiguously spaced by tp_frame_size bytes. If not, each
tp_block_size/tp_frame_size frames there will be a gap between 
the frames. This is because a frame cannot be spawn across two
blocks. 

At the beginning of each frame there is an status field (see 
struct tpacket_hdr). If this field is 0 means that the frame is ready
to be used for the kernel, If not, there is a frame the user can read 
and the following flags apply:

     from include/linux/if_packet.h

     #define TP_STATUS_COPY          2 
     #define TP_STATUS_LOSING        4 
     #define TP_STATUS_CSUMNOTREADY  8 


TP_STATUS_COPY        : This flag indicates that the frame (and associated
                        meta information) has been truncated because it's 
                        larger than tp_frame_size. This packet can be 
                        read entirely with recvfrom().
                        
                        In order to make this work it must to be
                        enabled previously with setsockopt() and 
                        the PACKET_COPY_THRESH option. 

                        The number of frames than can be buffered to 
                        be read with recvfrom is limited like a normal socket.
                        See the SO_RCVBUF option in the socket (7) man page.

TP_STATUS_LOSING      : indicates there were packet drops from last time 
                        statistics where checked with getsockopt() and
                        the PACKET_STATISTICS option.

TP_STATUS_CSUMNOTREADY: currently it's used for outgoing IP packets which 
                        it's checksum will be done in hardware. So while 
                        reading the packet we should not try to check the 
                        checksum. 

for convenience there are also the following defines:

     #define TP_STATUS_KERNEL        0
     #define TP_STATUS_USER          1

The kernel initializes all frames to TP_STATUS_KERNEL, when the kernel
receives a packet it puts in the buffer and updates the status with
at least the TP_STATUS_USER flag. Then the user can read the packet,
once the packet is read the user must zero the status field, so the kernel 
can use again that frame buffer.

The user can use poll (any other variant should apply too) to check if new
packets are in the ring:

    struct pollfd pfd;

    pfd.fd = fd;
    pfd.revents = 0;
    pfd.events = POLLIN|POLLRDNORM|POLLERR;

    if (status == TP_STATUS_KERNEL)
        retval = poll(&pfd, 1, timeout);

It doesn't incur in a race condition to first check the status value and 
then poll for frames.

--------------------------------------------------------------------------------
+ THANKS
--------------------------------------------------------------------------------
   
   Jesse Brandeburg, for fixing my grammathical/spelling errors