2 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``S IS''AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * Definitions of constants and the structures used by the netmap
31 * framework, for the part visible to both kernel and userspace.
32 * Detailed info on netmap is available with "man netmap" or at
34 * http://info.iet.unipi.it/~luigi/netmap/
36 * This API is also used to communicate with the VALE software switch
39 #ifndef _NET_NETMAP_H_
40 #define _NET_NETMAP_H_
42 #define NETMAP_API 10 /* current API version */
45 * Some fields should be cache-aligned to reduce contention.
46 * The alignment is architecture and OS dependent, but rather than
47 * digging into OS headers to find the exact value we use an estimate
48 * that should cover most architectures.
50 #define NM_CACHE_ALIGN 128
53 * --- Netmap data structures ---
55 * The userspace data structures used by netmap are shown below.
56 * They are allocated by the kernel and mmap()ed by userspace threads.
57 * Pointers are implemented as memory offsets or indexes,
58 * so that they can be easily dereferenced in kernel and userspace.
60 KERNEL (opaque, obviously)
62 ====================================================================
64 USERSPACE | struct netmap_ring
65 +---->+---------------+
67 struct netmap_if (nifp, 1 per fd) / | buf_ofs |
68 +---------------+ / | other fields |
69 | ni_tx_rings | / +===============+
70 | ni_rx_rings | / | buf_idx, len | slot[0]
72 | | / +---------------+
73 +===============+ / | buf_idx, len | slot[1]
74 | txring_ofs[0] | (rel.to nifp)--' | flags, ptr |
75 | txring_ofs[1] | +---------------+
76 (tx+1+extra_tx entries) (num_slots entries)
77 | txring_ofs[t] | | buf_idx, len | slot[n-1]
78 +---------------+ | flags, ptr |
79 | rxring_ofs[0] | +---------------+
81 (rx+1+extra_rx entries)
85 * For each "interface" (NIC, host stack, VALE switch port) attached to a
86 * file descriptor, the mmap()ed region contains a (logically readonly)
87 * struct netmap_if pointing to struct netmap_ring's.
88 * There is one netmap_ring per physical NIC ring, plus one tx/rx ring
89 * pair attached to the host stack (this pair is unused for VALE ports).
91 * All physical/host stack ports share the same memory region,
92 * so that zero-copy can be implemented between them.
93 * VALE switch ports instead have separate memory regions.
95 * The netmap_ring is the userspace-visible replica of the NIC ring.
96 * Each slot has the index of a buffer (MTU-sized and residing in the
97 * mmapped region), its length and some flags. An extra 64-bit pointer
98 * is provided for user-supplied buffers in the tx path.
100 * In user space, the buffer address is computed as
101 * (char *)ring + buf_ofs + index*NETMAP_BUF_SIZE
105 * struct netmap_slot is a buffer descriptor
108 uint32_t buf_idx; /* buffer index */
109 uint16_t len; /* length for this slot */
110 uint16_t flags; /* buf changed, etc. */
111 uint64_t ptr; /* pointer for indirect buffers */
115 * The following flags control how the slot is used
118 #define NS_BUF_CHANGED 0x0001 /* buf_idx changed */
120 * must be set whenever buf_idx is changed (as it might be
121 * necessary to recompute the physical address and mapping)
124 #define NS_REPORT 0x0002 /* ask the hardware to report results */
126 * Request notification when slot is used by the hardware.
127 * Normally transmit completions are handled lazily and
128 * may be unreported. This flag lets us know when a slot
129 * has been sent (e.g. to terminate the sender).
132 #define NS_FORWARD 0x0004 /* pass packet 'forward' */
134 * (Only for physical ports, rx rings with NR_FORWARD set).
135 * Slot released to the kernel (i.e. before ring->head) with
136 * this flag set are passed to the peer ring (host/NIC),
137 * thus restoring the host-NIC connection for these slots.
138 * This supports efficient traffic monitoring or firewalling.
141 #define NS_NO_LEARN 0x0008 /* disable bridge learning */
143 * On a VALE switch, do not 'learn' the source port for
147 #define NS_INDIRECT 0x0010 /* userspace buffer */
149 * (VALE tx rings only) data is in a userspace buffer,
150 * whose address is in the 'ptr' field in the slot.
153 #define NS_MOREFRAG 0x0020 /* packet has more fragments */
156 * Set on all but the last slot of a multi-segment packet.
157 * The 'len' field refers to the individual fragment.
160 #define NS_PORT_SHIFT 8
161 #define NS_PORT_MASK (0xff << NS_PORT_SHIFT)
163 * The high 8 bits of the flag, if not zero, indicate the
164 * destination port for the VALE switch, overriding
168 #define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff)
170 * (VALE rx rings only) the high 8 bits
171 * are the number of fragments.
178 * Netmap representation of a TX or RX ring (also known as "queue").
179 * This is a queue implemented as a fixed-size circular array.
180 * At the software level the important fields are: head, cur, tail.
184 * head first slot available for transmission.
185 * cur wakeup point. select() and poll() will unblock
186 * when 'tail' moves past 'cur'
187 * tail (readonly) first slot reserved to the kernel
189 * [head .. tail-1] can be used for new packets to send;
190 * 'head' and 'cur' must be incremented as slots are filled
191 * with new packets to be sent;
192 * 'cur' can be moved further ahead if we need more space
193 * for new transmissions.
197 * head first valid received packet
198 * cur wakeup point. select() and poll() will unblock
199 * when 'tail' moves past 'cur'
200 * tail (readonly) first slot reserved to the kernel
202 * [head .. tail-1] contain received packets;
203 * 'head' and 'cur' must be incremented as slots are consumed
204 * and can be returned to the kernel;
205 * 'cur' can be moved further ahead if we want to wait for
206 * new packets without returning the previous ones.
208 * DATA OWNERSHIP/LOCKING:
209 * The netmap_ring, and all slots and buffers in the range
210 * [head .. tail-1] are owned by the user program;
211 * the kernel only accesses them during a netmap system call
212 * and in the user thread context.
214 * Other slots and buffers are reserved for use by the kernel
218 * buf_ofs is meant to be used through macros.
219 * It contains the offset of the buffer region from this
222 const int64_t buf_ofs;
223 const uint32_t num_slots; /* number of slots in the ring. */
224 const uint32_t nr_buf_size;
225 const uint16_t ringid;
226 const uint16_t dir; /* 0: tx, 1: rx */
228 uint32_t head; /* (u) first user slot */
229 uint32_t cur; /* (u) wakeup point */
230 uint32_t tail; /* (k) first kernel slot */
234 struct timeval ts; /* (k) time of last *sync() */
236 /* opaque room for a mutex or similar object */
237 uint8_t sem[128] __attribute__((__aligned__(NM_CACHE_ALIGN)));
239 /* the slots follow. This struct has variable size */
240 struct netmap_slot slot[0]; /* array of slots. */
247 #define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */
249 * updates the 'ts' field on each netmap syscall. This saves
250 * saves a separate gettimeofday(), and is not much worse than
251 * software timestamps generated in the interrupt handler.
254 #define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */
256 * Enables the NS_FORWARD slot flag for the ring.
261 * Netmap representation of an interface and its queue(s).
262 * This is initialized by the kernel when binding a file
263 * descriptor to a port, and should be considered as readonly
264 * by user programs. The kernel never uses it.
266 * There is one netmap_if for each file descriptor on which we want
268 * select/poll operates on one or all pairs depending on the value of
269 * nmr_queueid passed on the ioctl.
272 char ni_name[IFNAMSIZ]; /* name of the interface. */
273 const uint32_t ni_version; /* API version, currently unused */
274 const uint32_t ni_flags; /* properties */
275 #define NI_PRIV_MEM 0x1 /* private memory region */
278 * The number of packet rings available in netmap mode.
279 * Physical NICs can have different numbers of tx and rx rings.
280 * Physical NICs also have a 'host' ring pair.
281 * Additionally, clients can request additional ring pairs to
282 * be used for internal communication.
284 const uint32_t ni_tx_rings; /* number of HW tx rings */
285 const uint32_t ni_rx_rings; /* number of HW rx rings */
287 const uint32_t ni_extra_tx_rings;
288 const uint32_t ni_extra_rx_rings;
290 * The following array contains the offset of each netmap ring
291 * from this structure, in the following order:
292 * NIC tx rings (ni_tx_rings); host tx ring (1); extra tx rings;
293 * NIC rx rings (ni_rx_rings); host tx ring (1); extra rx rings.
295 * The area is filled up by the kernel on NIOCREGIF,
296 * and then only read by userspace code.
298 const ssize_t ring_ofs[0];
304 * ioctl names and related fields
306 * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
307 * whose identity is set in NIOCREGIF through nr_ringid.
308 * These are non blocking and take no argument.
310 * NIOCGINFO takes a struct ifreq, the interface name is the input,
311 * the outputs are number of queues and number of descriptor
312 * for each queue (useful to set number of threads etc.).
313 * The info returned is only advisory and may change before
314 * the interface is bound to a file descriptor.
316 * NIOCREGIF takes an interface name within a struct nmre,
317 * and activates netmap mode on the interface (if possible).
319 * The argument to NIOCGINFO/NIOCREGIF overlays struct ifreq so we
320 * can pass it down to other NIC-related ioctls.
322 * The actual argument (struct nmreq) has a number of options to request
323 * different functions.
326 * The name of the port (em0, valeXXX:YYY, etc.)
327 * limited to IFNAMSIZ for backward compatibility.
329 * nr_version (in/out)
330 * Must match NETMAP_API as used in the kernel, error otherwise.
331 * Always returns the desired value on output.
333 * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings (in/out)
334 * On input, non-zero values may be used to reconfigure the port
335 * according to the requested values, but this is not guaranteed.
336 * On output the actual values in use are reported.
339 * Indicates how rings should be bound to the file descriptors.
340 * 0 (default) binds all physical rings
341 * NETMAP_HW_RING | ring number binds a single ring pair
342 * NETMAP_SW_RING binds only the host tx/rx rings
344 * NETMAP_NO_TX_POLL can be OR-ed to make select()/poll() push
345 * packets on tx rings only if POLLOUT is set.
346 * The default is to push any pending packet.
348 * NETMAP_PRIV_MEM is set on return for ports that use private
349 * memory regions and cannot use buffer swapping.
351 * nr_cmd (in) if non-zero indicates a special command:
352 * NETMAP_BDG_ATTACH and nr_name = vale*:ifname
353 * attaches the NIC to the switch; nr_ringid specifies
354 * which rings to use. Used by vale-ctl -a ...
355 * nr_arg1 = NETMAP_BDG_HOST also attaches the host port
356 * as in vale-ctl -h ...
358 * NETMAP_BDG_DETACH and nr_name = vale*:ifname
359 * disconnects a previously attached NIC.
360 * Used by vale-ctl -d ...
363 * list the configuration of VALE switches.
365 * NETMAP_BDG_OFFSET XXX ?
366 * Set the offset of data in packets. Used with VALE
367 * switches where the clients use the vhost header.
369 * nr_arg1, nr_arg2 (in/out) command specific
375 * struct nmreq overlays a struct ifreq
378 char nr_name[IFNAMSIZ];
379 uint32_t nr_version; /* API version */
380 uint32_t nr_offset; /* nifp offset in the shared region */
381 uint32_t nr_memsize; /* size of the shared region */
382 uint32_t nr_tx_slots; /* slots in tx rings */
383 uint32_t nr_rx_slots; /* slots in rx rings */
384 uint16_t nr_tx_rings; /* number of tx rings */
385 uint16_t nr_rx_rings; /* number of rx rings */
386 uint16_t nr_ringid; /* ring(s) we care about */
387 #define NETMAP_PRIV_MEM 0x8000 /* rings use private memory */
388 #define NETMAP_HW_RING 0x4000 /* low bits indicate one hw ring */
389 #define NETMAP_SW_RING 0x2000 /* process the sw ring */
390 #define NETMAP_NO_TX_POLL 0x1000 /* no automatic txsync on poll */
391 #define NETMAP_RING_MASK 0xfff /* the ring number */
394 #define NETMAP_BDG_ATTACH 1 /* attach the NIC */
395 #define NETMAP_BDG_DETACH 2 /* detach the NIC */
396 #define NETMAP_BDG_LOOKUP_REG 3 /* register lookup function */
397 #define NETMAP_BDG_LIST 4 /* get bridge's info */
398 #define NETMAP_BDG_OFFSET 5 /* set the port offset */
401 #define NETMAP_BDG_HOST 1 /* attach the host stack on ATTACH */
402 #define NETMAP_BDG_MAX_OFFSET 12
410 * FreeBSD uses the size value embedded in the _IOWR to determine
411 * how much to copy in/out. So we need it to match the actual
412 * data structure we pass. We put some spares in the structure
413 * to ease compatibility with other versions
415 #define NIOCGINFO _IOWR('i', 145, struct nmreq) /* return IF info */
416 #define NIOCREGIF _IOWR('i', 146, struct nmreq) /* interface register */
417 #define NIOCTXSYNC _IO('i', 148) /* sync tx queues */
418 #define NIOCRXSYNC _IO('i', 149) /* sync rx queues */
419 #endif /* !NIOCREGIF */
423 * Helper functions for kernel and userspace
427 * check if space is available in the ring.
430 nm_ring_empty(struct netmap_ring *ring)
432 return (ring->cur == ring->tail);
435 #endif /* _NET_NETMAP_H_ */