4 If you agree with it I will send two small patches to modify
5 kernel's configure help.
9 --------------------------------------------------------------------------------
11 --------------------------------------------------------------------------------
13 This file documents the CONFIG_PACKET_MMAP option available with the PACKET
14 socket interface on 2.4 and 2.6 kernels. This type of sockets is used for
15 capture network traffic with utilities like tcpdump or any other that uses
18 You can find the latest version of this document at
20 http://pusa.uv.es/~ulisses/packet_mmap/
22 Please send me your comments to
24 Ulisses Alonso CamarĂ³ <uaca@i.hate.spam.alumni.uv.es>
26 -------------------------------------------------------------------------------
28 --------------------------------------------------------------------------------
30 In Linux 2.4/2.6 if PACKET_MMAP is not enabled, the capture process is very
31 inefficient. It uses very limited buffers and requires one system call
32 to capture each packet, it requires two if you want to get packet's
33 timestamp (like libpcap always does).
35 In the other hand PACKET_MMAP is very efficient. PACKET_MMAP provides a size
36 configurable circular buffer mapped in user space. This way reading packets just
37 needs to wait for them, most of the time there is no need to issue a single
38 system call. By using a shared buffer between the kernel and the user
39 also has the benefit of minimizing packet copies.
41 It's fine to use PACKET_MMAP to improve the performance of the capture process,
42 but it isn't everything. At least, if you are capturing at high speeds (this
43 is relative to the cpu speed), you should check if the device driver of your
44 network interface card supports some sort of interrupt load mitigation or
45 (even better) if it supports NAPI, also make sure it is enabled.
47 --------------------------------------------------------------------------------
48 + How to use CONFIG_PACKET_MMAP
49 --------------------------------------------------------------------------------
51 From the user standpoint, you should use the higher level libpcap library, wich
52 is a de facto standard, portable across nearly all operating systems
55 Said that, at time of this writing, official libpcap 0.8.1 is out and doesn't include
56 support for PACKET_MMAP, and also probably the libpcap included in your distribution.
58 I'm aware of two implementations of PACKET_MMAP in libpcap:
60 http://pusa.uv.es/~ulisses/packet_mmap/ (by Simon Patarin, based on libpcap 0.6.2)
61 http://public.lanl.gov/cpw/ (by Phil Wood, based on lastest libpcap)
63 The rest of this document is intended for people who want to understand
64 the low level details or want to improve libpcap by including PACKET_MMAP
67 --------------------------------------------------------------------------------
68 + How to use CONFIG_PACKET_MMAP directly
69 --------------------------------------------------------------------------------
71 From the system calls stand point, the use of PACKET_MMAP involves
72 the following process:
75 [setup] socket() -------> creation of the capture socket
76 setsockopt() ---> allocation of the circular buffer (ring)
77 mmap() ---------> maping of the allocated buffer to the
80 [capture] poll() ---------> to wait for incoming packets
82 [shutdown] close() --------> destruction of the capture socket and
83 deallocation of all associated
87 socket creation and destruction is straight forward, and is done
88 the same way with or without PACKET_MMAP:
92 fd= socket(PF_PACKET, mode, htons(ETH_P_ALL))
94 where mode is SOCK_RAW for the raw interface were link level
95 information can be captured or SOCK_DGRAM for the cooked
96 interface where link level information capture is not
97 supported and a link level pseudo-header is provided
100 The destruction of the socket and all associated resources
101 is done by a simple call to close(fd).
103 Next I will describe PACKET_MMAP settings and it's constraints,
104 also the maping of the circular buffer in the user process and
105 the use of this buffer.
107 --------------------------------------------------------------------------------
108 + PACKET_MMAP settings
109 --------------------------------------------------------------------------------
112 To setup PACKET_MMAP from user level code is done with a call like
114 setsockopt(fd, SOL_PACKET, PACKET_RX_RING, (void *) &req, sizeof(req))
116 The most significant argument in the previous call is the req parameter,
117 this parameter must to have the following structure:
121 unsigned int tp_block_size; /* Minimal size of contiguous block */
122 unsigned int tp_block_nr; /* Number of blocks */
123 unsigned int tp_frame_size; /* Size of frame */
124 unsigned int tp_frame_nr; /* Total number of frames */
127 This structure is defined in /usr/include/linux/if_packet.h and establishes a
128 circular buffer (ring) of unswappable memory mapped in the capture process.
129 Being mapped in the capture process allows reading the captured frames and
130 related meta-information like timestamps without requiring a system call.
132 Captured frames are grouped in blocks. Each block is a physically contiguous
133 region of memory and holds tp_block_size/tp_frame_size frames. The total number
134 of blocks is tp_block_nr. Note that tp_frame_nr is a redundant parameter because
136 frames_per_block = tp_block_size/tp_frame_size
138 indeed, packet_set_ring checks that the following condition is true
140 frames_per_block * tp_block_nr == tp_frame_nr
143 Lets see an example, with the following values:
150 we will get the following buffer structure:
153 +---------+---------+ +---------+---------+
154 | frame 1 | frame 2 | | frame 3 | frame 4 |
155 +---------+---------+ +---------+---------+
158 +---------+---------+ +---------+---------+
159 | frame 5 | frame 6 | | frame 7 | frame 8 |
160 +---------+---------+ +---------+---------+
162 A frame can be of any size with the only condition it can fit in a block. A block
163 can only hold an integer number of frames, or in other words, a frame cannot
164 be spawn accross two blocks so there are some datails you have to take into
165 account when choosing the frame_size. See "Maping and use of the circular
169 --------------------------------------------------------------------------------
170 + PACKET_MMAP setting constraints
171 --------------------------------------------------------------------------------
173 In kernel versions prior to 2.4.26 (for the 2.4 branch) and 2.6.5 (2.6 branch),
174 the PACKET_MMAP buffer could hold only 32768 frames in a 32 bit architecture or
175 16384 in a 64 bit architecture. For information on these kernel versions
176 see http://pusa.uv.es/~ulisses/packet_mmap/packet_mmap.pre-2.4.26_2.6.5.txt
181 As stated earlier, each block is a contiguous physical region of memory. These
182 memory regions are allocated with calls to the __get_free_pages() function. As
183 the name indicates, this function allocates pages of memory, and the second
184 argument is "order" or a power of two number of pages, that is
185 (for PAGE_SIZE == 4096) order=0 ==> 4096 bytes, order=1 ==> 8192 bytes,
186 order=2 ==> 16384 bytes, etc. The maximum size of a
187 region allocated by __get_free_pages is determined by the MAX_ORDER macro. More
188 precisely the limit can be calculated as:
190 PAGE_SIZE << MAX_ORDER
192 In a i386 architecture PAGE_SIZE is 4096 bytes
193 In a 2.4/i386 kernel MAX_ORDER is 10
194 In a 2.6/i386 kernel MAX_ORDER is 11
196 So get_free_pages can allocate as much as 4MB or 8MB in a 2.4/2.6 kernel
197 respectively, with an i386 architecture.
199 User space programs can include /usr/include/sys/user.h and
200 /usr/include/linux/mmzone.h to get PAGE_SIZE MAX_ORDER declarations.
202 The pagesize can also be determined dynamically with the getpagesize (2)
209 To understand the constraints of PACKET_MMAP, we have to see the structure
210 used to hold the pointers to each block.
212 Currently, this structure is a dynamically allocated vector with kmalloc
213 called pg_vec, its size limits the number of blocks that can be allocated.
226 kmalloc allocates any number of bytes of phisically contiguous memory from
227 a pool of pre-determined sizes. This pool of memory is mantained by the slab
228 allocator wich is at the end the responsible for doing the allocation and
229 hence wich imposes the maximum memory that kmalloc can allocate.
231 In a 2.4/2.6 kernel and the i386 architecture, the limit is 131072 bytes. The
232 predetermined sizes that kmalloc uses can be checked in the "size-<bytes>"
233 entries of /proc/slabinfo
235 In a 32 bit architecture, pointers are 4 bytes long, so the total number of
236 pointers to blocks is
238 131072/4 = 32768 blocks
241 PACKET_MMAP buffer size calculator
242 ------------------------------------
246 <size-max> : is the maximum size of allocable with kmalloc (see /proc/slabinfo)
247 <pointer size>: depends on the architecture -- sizeof(void *)
248 <page size> : depends on the architecture -- PAGE_SIZE or getpagesize (2)
249 <max-order> : is the value defined with MAX_ORDER
250 <frame size> : it's an upper bound of frame's capture size (more on this later)
252 from these definitions we will derive
254 <block number> = <size-max>/<pointer size>
255 <block size> = <pagesize> << <max-order>
257 so, the max buffer size is
259 <block number> * <block size>
261 and, the number of frames be
263 <block number> * <block size> / <frame size>
265 Suposse the following parameters, wich apply for 2.6 kernel and an
268 <size-max> = 131072 bytes
269 <pointer size> = 4 bytes
270 <pagesize> = 4096 bytes
273 and a value for <frame size> of 2048 byteas. These parameters will yield
275 <block number> = 131072/4 = 32768 blocks
276 <block size> = 4096 << 11 = 8 MiB.
278 and hence the buffer will have a 262144 MiB size. So it can hold
279 262144 MiB / 2048 bytes = 134217728 frames
282 Actually, this buffer size is not possible with an i386 architecture.
283 Remember that the memory is allocated in kernel space, in the case of
284 an i386 kernel's memory size is limited to 1GiB.
286 All memory allocations are not freed until the socket is closed. The memory
287 allocations are done with GFP_KERNEL priority, this basically means that
288 the allocation can wait and swap other process' memory in order to allocate
289 the nececessary memory, so normally limits can be reached.
294 If you check the source code you will see that what I draw here as a frame
295 is not only the link level frame. At the begining of each frame there is a
296 header called struct tpacket_hdr used in PACKET_MMAP to hold link level's frame
297 meta information like timestamp. So what we draw here a frame it's really
298 the following (from include/linux/if_packet.h):
303 - Start. Frame must be aligned to TPACKET_ALIGNMENT=16
305 - pad to TPACKET_ALIGNMENT=16
307 - Gap, chosen so that packet data (Start+tp_net) alignes to
309 - Start+tp_mac: [ Optional MAC header ]
310 - Start+tp_net: Packet data, aligned to TPACKET_ALIGNMENT=16.
311 - Pad to align to TPACKET_ALIGNMENT=16
315 The following are conditions that are checked in packet_set_ring
317 tp_block_size must be a multiple of PAGE_SIZE (1)
318 tp_frame_size must be greater than TPACKET_HDRLEN (obvious)
319 tp_frame_size must be a multiple of TPACKET_ALIGNMENT
320 tp_frame_nr must be exactly frames_per_block*tp_block_nr
322 Note that tp_block_size should be choosed to be a power of two or there will
323 be a waste of memory.
325 --------------------------------------------------------------------------------
326 + Maping and use of the circular buffer (ring)
327 --------------------------------------------------------------------------------
329 The maping of the buffer in the user process is done with the conventional
330 mmap function. Even the circular buffer is compound of several physically
331 discontiguous blocks of memory, they are contiguous to the user space, hence
332 just one call to mmap is needed:
334 mmap(0, size, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
336 If tp_frame_size is a divisor of tp_block_size frames will be
337 contiguosly spaced by tp_frame_size bytes. If not, each
338 tp_block_size/tp_frame_size frames there will be a gap between
339 the frames. This is because a frame cannot be spawn across two
342 At the beginning of each frame there is an status field (see
343 struct tpacket_hdr). If this field is 0 means that the frame is ready
344 to be used for the kernel, If not, there is a frame the user can read
345 and the following flags apply:
347 from include/linux/if_packet.h
349 #define TP_STATUS_COPY 2
350 #define TP_STATUS_LOSING 4
351 #define TP_STATUS_CSUMNOTREADY 8
354 TP_STATUS_COPY : This flag indicates that the frame (and associated
355 meta information) has been truncated because it's
356 larger than tp_frame_size. This packet can be
357 read entirely with recvfrom().
359 In order to make this work it must to be
360 enabled previously with setsockopt() and
361 the PACKET_COPY_THRESH option.
363 The number of frames than can be buffered to
364 be read with recvfrom is limited like a normal socket.
365 See the SO_RCVBUF option in the socket (7) man page.
367 TP_STATUS_LOSING : indicates there were packet drops from last time
368 statistics where checked with getsockopt() and
369 the PACKET_STATISTICS option.
371 TP_STATUS_CSUMNOTREADY: currently it's used for outgoing IP packets wich
372 it's checksum will be done in hardware. So while
373 reading the packet we should not try to check the
376 for convenience there are also the following defines:
378 #define TP_STATUS_KERNEL 0
379 #define TP_STATUS_USER 1
381 The kernel initializes all frames to TP_STATUS_KERNEL, when the kernel
382 receives a packet it puts in the buffer and updates the status with
383 at least the TP_STATUS_USER flag. Then the user can read the packet,
384 once the packet is read the user must zero the status field, so the kernel
385 can use again that frame buffer.
387 The user can use poll (any other variant should apply too) to check if new
388 packets are in the ring:
394 pfd.events = POLLIN|POLLRDNORM|POLLERR;
396 if (status == TP_STATUS_KERNEL)
397 retval = poll(&pfd, 1, timeout);
399 It doesn't incur in a race condition to first check the status value and
400 then poll for frames.
402 --------------------------------------------------------------------------------
404 --------------------------------------------------------------------------------
406 Jesse Brandeburg, for fixing my grammathical/spelling errors
410 To unsubscribe from this list: send the line "unsubscribe linux-net" in
411 the body of a message to majordomo@vger.kernel.org
412 More majordomo info at http://vger.kernel.org/majordomo-info.html