xref: /freebsd/sys/net/netmap.h (revision f4b37ed0f8b307b1f3f0f630ca725d68f1dff30d)
1 /*
2  * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  *
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.
13  *
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
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * $FreeBSD$
29  *
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
33  *
34  *	http://info.iet.unipi.it/~luigi/netmap/
35  *
36  * This API is also used to communicate with the VALE software switch
37  */
38 
39 #ifndef _NET_NETMAP_H_
40 #define _NET_NETMAP_H_
41 
42 #define	NETMAP_API	11		/* current API version */
43 
44 #define	NETMAP_MIN_API	11		/* min and max versions accepted */
45 #define	NETMAP_MAX_API	15
46 /*
47  * Some fields should be cache-aligned to reduce contention.
48  * The alignment is architecture and OS dependent, but rather than
49  * digging into OS headers to find the exact value we use an estimate
50  * that should cover most architectures.
51  */
52 #define NM_CACHE_ALIGN	128
53 
54 /*
55  * --- Netmap data structures ---
56  *
57  * The userspace data structures used by netmap are shown below.
58  * They are allocated by the kernel and mmap()ed by userspace threads.
59  * Pointers are implemented as memory offsets or indexes,
60  * so that they can be easily dereferenced in kernel and userspace.
61 
62    KERNEL (opaque, obviously)
63 
64   ====================================================================
65                                          |
66    USERSPACE                             |      struct netmap_ring
67                                          +---->+---------------+
68                                              / | head,cur,tail |
69    struct netmap_if (nifp, 1 per fd)        /  | buf_ofs       |
70     +---------------+                      /   | other fields  |
71     | ni_tx_rings   |                     /    +===============+
72     | ni_rx_rings   |                    /     | buf_idx, len  | slot[0]
73     |               |                   /      | flags, ptr    |
74     |               |                  /       +---------------+
75     +===============+                 /        | buf_idx, len  | slot[1]
76     | txring_ofs[0] | (rel.to nifp)--'         | flags, ptr    |
77     | txring_ofs[1] |                          +---------------+
78      (tx+1 entries)                           (num_slots entries)
79     | txring_ofs[t] |                          | buf_idx, len  | slot[n-1]
80     +---------------+                          | flags, ptr    |
81     | rxring_ofs[0] |                          +---------------+
82     | rxring_ofs[1] |
83      (rx+1 entries)
84     | rxring_ofs[r] |
85     +---------------+
86 
87  * For each "interface" (NIC, host stack, PIPE, VALE switch port) bound to
88  * a file descriptor, the mmap()ed region contains a (logically readonly)
89  * struct netmap_if pointing to struct netmap_ring's.
90  *
91  * There is one netmap_ring per physical NIC ring, plus one tx/rx ring
92  * pair attached to the host stack (this pair is unused for non-NIC ports).
93  *
94  * All physical/host stack ports share the same memory region,
95  * so that zero-copy can be implemented between them.
96  * VALE switch ports instead have separate memory regions.
97  *
98  * The netmap_ring is the userspace-visible replica of the NIC ring.
99  * Each slot has the index of a buffer (MTU-sized and residing in the
100  * mmapped region), its length and some flags. An extra 64-bit pointer
101  * is provided for user-supplied buffers in the tx path.
102  *
103  * In user space, the buffer address is computed as
104  *	(char *)ring + buf_ofs + index * NETMAP_BUF_SIZE
105  *
106  * Added in NETMAP_API 11:
107  *
108  * + NIOCREGIF can request the allocation of extra spare buffers from
109  *   the same memory pool. The desired number of buffers must be in
110  *   nr_arg3. The ioctl may return fewer buffers, depending on memory
111  *   availability. nr_arg3 will return the actual value, and, once
112  *   mapped, nifp->ni_bufs_head will be the index of the first buffer.
113  *
114  *   The buffers are linked to each other using the first uint32_t
115  *   as the index. On close, ni_bufs_head must point to the list of
116  *   buffers to be released.
117  *
118  * + NIOCREGIF can request space for extra rings (and buffers)
119  *   allocated in the same memory space. The number of extra rings
120  *   is in nr_arg1, and is advisory. This is a no-op on NICs where
121  *   the size of the memory space is fixed.
122  *
123  * + NIOCREGIF can attach to PIPE rings sharing the same memory
124  *   space with a parent device. The ifname indicates the parent device,
125  *   which must already exist. Flags in nr_flags indicate if we want to
126  *   bind the master or slave side, the index (from nr_ringid)
127  *   is just a cookie and does not need to be sequential.
128  *
129  * + NIOCREGIF can also attach to 'monitor' rings that replicate
130  *   the content of specific rings, also from the same memory space.
131  *
132  *   Extra flags in nr_flags support the above functions.
133  *   Application libraries may use the following naming scheme:
134  *	netmap:foo			all NIC ring pairs
135  *	netmap:foo^			only host ring pair
136  *	netmap:foo+			all NIC ring + host ring pairs
137  *	netmap:foo-k			the k-th NIC ring pair
138  *	netmap:foo{k			PIPE ring pair k, master side
139  *	netmap:foo}k			PIPE ring pair k, slave side
140  */
141 
142 /*
143  * struct netmap_slot is a buffer descriptor
144  */
145 struct netmap_slot {
146 	uint32_t buf_idx;	/* buffer index */
147 	uint16_t len;		/* length for this slot */
148 	uint16_t flags;		/* buf changed, etc. */
149 	uint64_t ptr;		/* pointer for indirect buffers */
150 };
151 
152 /*
153  * The following flags control how the slot is used
154  */
155 
156 #define	NS_BUF_CHANGED	0x0001	/* buf_idx changed */
157 	/*
158 	 * must be set whenever buf_idx is changed (as it might be
159 	 * necessary to recompute the physical address and mapping)
160 	 *
161 	 * It is also set by the kernel whenever the buf_idx is
162 	 * changed internally (e.g., by pipes). Applications may
163 	 * use this information to know when they can reuse the
164 	 * contents of previously prepared buffers.
165 	 */
166 
167 #define	NS_REPORT	0x0002	/* ask the hardware to report results */
168 	/*
169 	 * Request notification when slot is used by the hardware.
170 	 * Normally transmit completions are handled lazily and
171 	 * may be unreported. This flag lets us know when a slot
172 	 * has been sent (e.g. to terminate the sender).
173 	 */
174 
175 #define	NS_FORWARD	0x0004	/* pass packet 'forward' */
176 	/*
177 	 * (Only for physical ports, rx rings with NR_FORWARD set).
178 	 * Slot released to the kernel (i.e. before ring->head) with
179 	 * this flag set are passed to the peer ring (host/NIC),
180 	 * thus restoring the host-NIC connection for these slots.
181 	 * This supports efficient traffic monitoring or firewalling.
182 	 */
183 
184 #define	NS_NO_LEARN	0x0008	/* disable bridge learning */
185  	/*
186 	 * On a VALE switch, do not 'learn' the source port for
187  	 * this buffer.
188 	 */
189 
190 #define	NS_INDIRECT	0x0010	/* userspace buffer */
191  	/*
192 	 * (VALE tx rings only) data is in a userspace buffer,
193 	 * whose address is in the 'ptr' field in the slot.
194 	 */
195 
196 #define	NS_MOREFRAG	0x0020	/* packet has more fragments */
197  	/*
198 	 * (VALE ports only)
199 	 * Set on all but the last slot of a multi-segment packet.
200 	 * The 'len' field refers to the individual fragment.
201 	 */
202 
203 #define	NS_PORT_SHIFT	8
204 #define	NS_PORT_MASK	(0xff << NS_PORT_SHIFT)
205 	/*
206  	 * The high 8 bits of the flag, if not zero, indicate the
207 	 * destination port for the VALE switch, overriding
208  	 * the lookup table.
209  	 */
210 
211 #define	NS_RFRAGS(_slot)	( ((_slot)->flags >> 8) & 0xff)
212 	/*
213 	 * (VALE rx rings only) the high 8 bits
214 	 *  are the number of fragments.
215 	 */
216 
217 
218 /*
219  * struct netmap_ring
220  *
221  * Netmap representation of a TX or RX ring (also known as "queue").
222  * This is a queue implemented as a fixed-size circular array.
223  * At the software level the important fields are: head, cur, tail.
224  *
225  * In TX rings:
226  *
227  *	head	first slot available for transmission.
228  *	cur	wakeup point. select() and poll() will unblock
229  *		when 'tail' moves past 'cur'
230  *	tail	(readonly) first slot reserved to the kernel
231  *
232  *	[head .. tail-1] can be used for new packets to send;
233  *	'head' and 'cur' must be incremented as slots are filled
234  *	    with new packets to be sent;
235  *	'cur' can be moved further ahead if we need more space
236  *	for new transmissions. XXX todo (2014-03-12)
237  *
238  * In RX rings:
239  *
240  *	head	first valid received packet
241  *	cur	wakeup point. select() and poll() will unblock
242  *		when 'tail' moves past 'cur'
243  *	tail	(readonly) first slot reserved to the kernel
244  *
245  *	[head .. tail-1] contain received packets;
246  *	'head' and 'cur' must be incremented as slots are consumed
247  *		and can be returned to the kernel;
248  *	'cur' can be moved further ahead if we want to wait for
249  *		new packets without returning the previous ones.
250  *
251  * DATA OWNERSHIP/LOCKING:
252  *	The netmap_ring, and all slots and buffers in the range
253  *	[head .. tail-1] are owned by the user program;
254  *	the kernel only accesses them during a netmap system call
255  *	and in the user thread context.
256  *
257  *	Other slots and buffers are reserved for use by the kernel
258  */
259 struct netmap_ring {
260 	/*
261 	 * buf_ofs is meant to be used through macros.
262 	 * It contains the offset of the buffer region from this
263 	 * descriptor.
264 	 */
265 	const int64_t	buf_ofs;
266 	const uint32_t	num_slots;	/* number of slots in the ring. */
267 	const uint32_t	nr_buf_size;
268 	const uint16_t	ringid;
269 	const uint16_t	dir;		/* 0: tx, 1: rx */
270 
271 	uint32_t        head;		/* (u) first user slot */
272 	uint32_t        cur;		/* (u) wakeup point */
273 	uint32_t	tail;		/* (k) first kernel slot */
274 
275 	uint32_t	flags;
276 
277 	struct timeval	ts;		/* (k) time of last *sync() */
278 
279 	/* opaque room for a mutex or similar object */
280 	uint8_t		sem[128] __attribute__((__aligned__(NM_CACHE_ALIGN)));
281 
282 	/* the slots follow. This struct has variable size */
283 	struct netmap_slot slot[0];	/* array of slots. */
284 };
285 
286 
287 /*
288  * RING FLAGS
289  */
290 #define	NR_TIMESTAMP	0x0002		/* set timestamp on *sync() */
291 	/*
292 	 * updates the 'ts' field on each netmap syscall. This saves
293 	 * saves a separate gettimeofday(), and is not much worse than
294 	 * software timestamps generated in the interrupt handler.
295 	 */
296 
297 #define	NR_FORWARD	0x0004		/* enable NS_FORWARD for ring */
298  	/*
299 	 * Enables the NS_FORWARD slot flag for the ring.
300 	 */
301 
302 
303 /*
304  * Netmap representation of an interface and its queue(s).
305  * This is initialized by the kernel when binding a file
306  * descriptor to a port, and should be considered as readonly
307  * by user programs. The kernel never uses it.
308  *
309  * There is one netmap_if for each file descriptor on which we want
310  * to select/poll.
311  * select/poll operates on one or all pairs depending on the value of
312  * nmr_queueid passed on the ioctl.
313  */
314 struct netmap_if {
315 	char		ni_name[IFNAMSIZ]; /* name of the interface. */
316 	const uint32_t	ni_version;	/* API version, currently unused */
317 	const uint32_t	ni_flags;	/* properties */
318 #define	NI_PRIV_MEM	0x1		/* private memory region */
319 
320 	/*
321 	 * The number of packet rings available in netmap mode.
322 	 * Physical NICs can have different numbers of tx and rx rings.
323 	 * Physical NICs also have a 'host' ring pair.
324 	 * Additionally, clients can request additional ring pairs to
325 	 * be used for internal communication.
326 	 */
327 	const uint32_t	ni_tx_rings;	/* number of HW tx rings */
328 	const uint32_t	ni_rx_rings;	/* number of HW rx rings */
329 
330 	uint32_t	ni_bufs_head;	/* head index for extra bufs */
331 	uint32_t	ni_spare1[5];
332 	/*
333 	 * The following array contains the offset of each netmap ring
334 	 * from this structure, in the following order:
335 	 * NIC tx rings (ni_tx_rings); host tx ring (1); extra tx rings;
336 	 * NIC rx rings (ni_rx_rings); host tx ring (1); extra rx rings.
337 	 *
338 	 * The area is filled up by the kernel on NIOCREGIF,
339 	 * and then only read by userspace code.
340 	 */
341 	const ssize_t	ring_ofs[0];
342 };
343 
344 
345 #ifndef NIOCREGIF
346 /*
347  * ioctl names and related fields
348  *
349  * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
350  *	whose identity is set in NIOCREGIF through nr_ringid.
351  *	These are non blocking and take no argument.
352  *
353  * NIOCGINFO takes a struct ifreq, the interface name is the input,
354  *	the outputs are number of queues and number of descriptor
355  *	for each queue (useful to set number of threads etc.).
356  *	The info returned is only advisory and may change before
357  *	the interface is bound to a file descriptor.
358  *
359  * NIOCREGIF takes an interface name within a struct nmre,
360  *	and activates netmap mode on the interface (if possible).
361  *
362  * The argument to NIOCGINFO/NIOCREGIF overlays struct ifreq so we
363  * can pass it down to other NIC-related ioctls.
364  *
365  * The actual argument (struct nmreq) has a number of options to request
366  * different functions.
367  * The following are used in NIOCREGIF when nr_cmd == 0:
368  *
369  * nr_name	(in)
370  *	The name of the port (em0, valeXXX:YYY, etc.)
371  *	limited to IFNAMSIZ for backward compatibility.
372  *
373  * nr_version	(in/out)
374  *	Must match NETMAP_API as used in the kernel, error otherwise.
375  *	Always returns the desired value on output.
376  *
377  * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings (in/out)
378  *	On input, non-zero values may be used to reconfigure the port
379  *	according to the requested values, but this is not guaranteed.
380  *	On output the actual values in use are reported.
381  *
382  * nr_ringid (in)
383  *	Indicates how rings should be bound to the file descriptors.
384  *	If nr_flags != 0, then the low bits (in NETMAP_RING_MASK)
385  *	are used to indicate the ring number, and nr_flags specifies
386  *	the actual rings to bind. NETMAP_NO_TX_POLL is unaffected.
387  *
388  *	NOTE: THE FOLLOWING (nr_flags == 0) IS DEPRECATED:
389  *	If nr_flags == 0, NETMAP_HW_RING and NETMAP_SW_RING control
390  *	the binding as follows:
391  *	0 (default)			binds all physical rings
392  *	NETMAP_HW_RING | ring number	binds a single ring pair
393  *	NETMAP_SW_RING			binds only the host tx/rx rings
394  *
395  *	NETMAP_NO_TX_POLL can be OR-ed to make select()/poll() push
396  *		packets on tx rings only if POLLOUT is set.
397  *		The default is to push any pending packet.
398  *
399  *	NETMAP_DO_RX_POLL can be OR-ed to make select()/poll() release
400  *		packets on rx rings also when POLLIN is NOT set.
401  *		The default is to touch the rx ring only with POLLIN.
402  *		Note that this is the opposite of TX because it
403  *		reflects the common usage.
404  *
405  *	NOTE: NETMAP_PRIV_MEM IS DEPRECATED, use nr_arg2 instead.
406  *	NETMAP_PRIV_MEM is set on return for ports that do not use
407  *		the global memory allocator.
408  *		This information is not significant and applications
409  *		should look at the region id in nr_arg2
410  *
411  * nr_flags	is the recommended mode to indicate which rings should
412  *		be bound to a file descriptor. Values are NR_REG_*
413  *
414  * nr_arg1 (in)	The number of extra rings to be reserved.
415  *		Especially when allocating a VALE port the system only
416  *		allocates the amount of memory needed for the port.
417  *		If more shared memory rings are desired (e.g. for pipes),
418  *		the first invocation for the same basename/allocator
419  *		should specify a suitable number. Memory cannot be
420  *		extended after the first allocation without closing
421  *		all ports on the same region.
422  *
423  * nr_arg2 (in/out) The identity of the memory region used.
424  *		On input, 0 means the system decides autonomously,
425  *		other values may try to select a specific region.
426  *		On return the actual value is reported.
427  *		Region '1' is the global allocator, normally shared
428  *		by all interfaces. Other values are private regions.
429  *		If two ports the same region zero-copy is possible.
430  *
431  * nr_arg3 (in/out)	number of extra buffers to be allocated.
432  *
433  *
434  *
435  * nr_cmd (in)	if non-zero indicates a special command:
436  *	NETMAP_BDG_ATTACH	 and nr_name = vale*:ifname
437  *		attaches the NIC to the switch; nr_ringid specifies
438  *		which rings to use. Used by vale-ctl -a ...
439  *	    nr_arg1 = NETMAP_BDG_HOST also attaches the host port
440  *		as in vale-ctl -h ...
441  *
442  *	NETMAP_BDG_DETACH	and nr_name = vale*:ifname
443  *		disconnects a previously attached NIC.
444  *		Used by vale-ctl -d ...
445  *
446  *	NETMAP_BDG_LIST
447  *		list the configuration of VALE switches.
448  *
449  *	NETMAP_BDG_VNET_HDR
450  *		Set the virtio-net header length used by the client
451  *		of a VALE switch port.
452  *
453  *	NETMAP_BDG_NEWIF
454  *		create a persistent VALE port with name nr_name.
455  *		Used by vale-ctl -n ...
456  *
457  *	NETMAP_BDG_DELIF
458  *		delete a persistent VALE port. Used by vale-ctl -d ...
459  *
460  * nr_arg1, nr_arg2, nr_arg3  (in/out)		command specific
461  *
462  *
463  *
464  */
465 
466 
467 /*
468  * struct nmreq overlays a struct ifreq (just the name)
469  */
470 struct nmreq {
471 	char		nr_name[IFNAMSIZ];
472 	uint32_t	nr_version;	/* API version */
473 	uint32_t	nr_offset;	/* nifp offset in the shared region */
474 	uint32_t	nr_memsize;	/* size of the shared region */
475 	uint32_t	nr_tx_slots;	/* slots in tx rings */
476 	uint32_t	nr_rx_slots;	/* slots in rx rings */
477 	uint16_t	nr_tx_rings;	/* number of tx rings */
478 	uint16_t	nr_rx_rings;	/* number of rx rings */
479 
480 	uint16_t	nr_ringid;	/* ring(s) we care about */
481 #define NETMAP_HW_RING		0x4000	/* single NIC ring pair */
482 #define NETMAP_SW_RING		0x2000	/* only host ring pair */
483 
484 #define NETMAP_RING_MASK	0x0fff	/* the ring number */
485 
486 #define NETMAP_NO_TX_POLL	0x1000	/* no automatic txsync on poll */
487 
488 #define NETMAP_DO_RX_POLL	0x8000	/* DO automatic rxsync on poll */
489 
490 	uint16_t	nr_cmd;
491 #define NETMAP_BDG_ATTACH	1	/* attach the NIC */
492 #define NETMAP_BDG_DETACH	2	/* detach the NIC */
493 #define NETMAP_BDG_REGOPS	3	/* register bridge callbacks */
494 #define NETMAP_BDG_LIST		4	/* get bridge's info */
495 #define NETMAP_BDG_VNET_HDR     5       /* set the port virtio-net-hdr length */
496 #define NETMAP_BDG_OFFSET	NETMAP_BDG_VNET_HDR	/* deprecated alias */
497 #define NETMAP_BDG_NEWIF	6	/* create a virtual port */
498 #define NETMAP_BDG_DELIF	7	/* destroy a virtual port */
499 	uint16_t	nr_arg1;	/* reserve extra rings in NIOCREGIF */
500 #define NETMAP_BDG_HOST		1	/* attach the host stack on ATTACH */
501 
502 	uint16_t	nr_arg2;
503 	uint32_t	nr_arg3;	/* req. extra buffers in NIOCREGIF */
504 	uint32_t	nr_flags;
505 	/* various modes, extends nr_ringid */
506 	uint32_t	spare2[1];
507 };
508 
509 #define NR_REG_MASK		0xf /* values for nr_flags */
510 enum {	NR_REG_DEFAULT	= 0,	/* backward compat, should not be used. */
511 	NR_REG_ALL_NIC	= 1,
512 	NR_REG_SW	= 2,
513 	NR_REG_NIC_SW	= 3,
514 	NR_REG_ONE_NIC	= 4,
515 	NR_REG_PIPE_MASTER = 5,
516 	NR_REG_PIPE_SLAVE = 6,
517 };
518 /* monitor uses the NR_REG to select the rings to monitor */
519 #define NR_MONITOR_TX	0x100
520 #define NR_MONITOR_RX	0x200
521 #define NR_ZCOPY_MON	0x400
522 /* request exclusive access to the selected rings */
523 #define NR_EXCLUSIVE	0x800
524 
525 
526 /*
527  * FreeBSD uses the size value embedded in the _IOWR to determine
528  * how much to copy in/out. So we need it to match the actual
529  * data structure we pass. We put some spares in the structure
530  * to ease compatibility with other versions
531  */
532 #define NIOCGINFO	_IOWR('i', 145, struct nmreq) /* return IF info */
533 #define NIOCREGIF	_IOWR('i', 146, struct nmreq) /* interface register */
534 #define NIOCTXSYNC	_IO('i', 148) /* sync tx queues */
535 #define NIOCRXSYNC	_IO('i', 149) /* sync rx queues */
536 #define NIOCCONFIG	_IOWR('i',150, struct nm_ifreq) /* for ext. modules */
537 #endif /* !NIOCREGIF */
538 
539 
540 /*
541  * Helper functions for kernel and userspace
542  */
543 
544 /*
545  * check if space is available in the ring.
546  */
547 static inline int
548 nm_ring_empty(struct netmap_ring *ring)
549 {
550 	return (ring->cur == ring->tail);
551 }
552 
553 /*
554  * Opaque structure that is passed to an external kernel
555  * module via ioctl(fd, NIOCCONFIG, req) for a user-owned
556  * bridge port (at this point ephemeral VALE interface).
557  */
558 #define NM_IFRDATA_LEN 256
559 struct nm_ifreq {
560 	char nifr_name[IFNAMSIZ];
561 	char data[NM_IFRDATA_LEN];
562 };
563 
564 #endif /* _NET_NETMAP_H_ */
565