xref: /freebsd/sys/net/netmap.h (revision b9f654b163bce26de79705e77b872427c9f2afa1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  *
10  *   1. Redistributions of source code must retain the above copyright
11  *      notice, this list of conditions and the following disclaimer.
12  *   2. Redistributions in binary form must reproduce the above copyright
13  *      notice, this list of conditions and the following disclaimer in the
14  *      documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``S IS''AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * $FreeBSD$
31  *
32  * Definitions of constants and the structures used by the netmap
33  * framework, for the part visible to both kernel and userspace.
34  * Detailed info on netmap is available with "man netmap" or at
35  *
36  *	http://info.iet.unipi.it/~luigi/netmap/
37  *
38  * This API is also used to communicate with the VALE software switch
39  */
40 
41 #ifndef _NET_NETMAP_H_
42 #define _NET_NETMAP_H_
43 
44 #define	NETMAP_API	14		/* current API version */
45 
46 #define	NETMAP_MIN_API	14		/* min and max versions accepted */
47 #define	NETMAP_MAX_API	15
48 /*
49  * Some fields should be cache-aligned to reduce contention.
50  * The alignment is architecture and OS dependent, but rather than
51  * digging into OS headers to find the exact value we use an estimate
52  * that should cover most architectures.
53  */
54 #define NM_CACHE_ALIGN	128
55 
56 /*
57  * --- Netmap data structures ---
58  *
59  * The userspace data structures used by netmap are shown below.
60  * They are allocated by the kernel and mmap()ed by userspace threads.
61  * Pointers are implemented as memory offsets or indexes,
62  * so that they can be easily dereferenced in kernel and userspace.
63 
64    KERNEL (opaque, obviously)
65 
66   ====================================================================
67                                           |
68    USERSPACE                              |      struct netmap_ring
69                                           +---->+---------------+
70                                               / | head,cur,tail |
71    struct netmap_if (nifp, 1 per fd)         /  | buf_ofs       |
72     +----------------+                      /   | other fields  |
73     | ni_tx_rings    |                     /    +===============+
74     | ni_rx_rings    |                    /     | buf_idx, len  | slot[0]
75     |                |                   /      | flags, ptr    |
76     |                |                  /       +---------------+
77     +================+                 /        | buf_idx, len  | slot[1]
78     | txring_ofs[0]  | (rel.to nifp)--'         | flags, ptr    |
79     | txring_ofs[1]  |                          +---------------+
80      (tx+htx entries)                           (num_slots entries)
81     | txring_ofs[t]  |                          | buf_idx, len  | slot[n-1]
82     +----------------+                          | flags, ptr    |
83     | rxring_ofs[0]  |                          +---------------+
84     | rxring_ofs[1]  |
85      (rx+hrx entries)
86     | rxring_ofs[r]  |
87     +----------------+
88 
89  * For each "interface" (NIC, host stack, PIPE, VALE switch port) bound to
90  * a file descriptor, the mmap()ed region contains a (logically readonly)
91  * struct netmap_if pointing to struct netmap_ring's.
92  *
93  * There is one netmap_ring per physical NIC ring, plus at least one tx/rx ring
94  * pair attached to the host stack (these pairs are unused for non-NIC ports).
95  *
96  * All physical/host stack ports share the same memory region,
97  * so that zero-copy can be implemented between them.
98  * VALE switch ports instead have separate memory regions.
99  *
100  * The netmap_ring is the userspace-visible replica of the NIC ring.
101  * Each slot has the index of a buffer (MTU-sized and residing in the
102  * mmapped region), its length and some flags. An extra 64-bit pointer
103  * is provided for user-supplied buffers in the tx path.
104  *
105  * In user space, the buffer address is computed as
106  *	(char *)ring + buf_ofs + index * NETMAP_BUF_SIZE
107  *
108  * Added in NETMAP_API 11:
109  *
110  * + NIOCREGIF can request the allocation of extra spare buffers from
111  *   the same memory pool. The desired number of buffers must be in
112  *   nr_arg3. The ioctl may return fewer buffers, depending on memory
113  *   availability. nr_arg3 will return the actual value, and, once
114  *   mapped, nifp->ni_bufs_head will be the index of the first buffer.
115  *
116  *   The buffers are linked to each other using the first uint32_t
117  *   as the index. On close, ni_bufs_head must point to the list of
118  *   buffers to be released.
119  *
120  * + NIOCREGIF can attach to PIPE rings sharing the same memory
121  *   space with a parent device. The ifname indicates the parent device,
122  *   which must already exist. Flags in nr_flags indicate if we want to
123  *   bind the master or slave side, the index (from nr_ringid)
124  *   is just a cookie and does not need to be sequential.
125  *
126  * + NIOCREGIF can also attach to 'monitor' rings that replicate
127  *   the content of specific rings, also from the same memory space.
128  *
129  *   Extra flags in nr_flags support the above functions.
130  *   Application libraries may use the following naming scheme:
131  *	netmap:foo			all NIC rings pairs
132  *	netmap:foo^			only host rings pairs
133  *	netmap:foo^k			the k-th host rings pair
134  *	netmap:foo+			all NIC rings + host rings pairs
135  *	netmap:foo-k			the k-th NIC rings pair
136  *	netmap:foo{k			PIPE rings pair k, master side
137  *	netmap:foo}k			PIPE rings pair k, slave side
138  *
139  * Some notes about host rings:
140  *
141  * + The RX host rings are used to store those packets that the host network
142  *   stack is trying to transmit through a NIC queue, but only if that queue
143  *   is currently in netmap mode. Netmap will not intercept host stack mbufs
144  *   designated to NIC queues that are not in netmap mode. As a consequence,
145  *   registering a netmap port with netmap:foo^ is not enough to intercept
146  *   mbufs in the RX host rings; the netmap port should be registered with
147  *   netmap:foo*, or another registration should be done to open at least a
148  *   NIC TX queue in netmap mode.
149  *
150  * + Netmap is not currently able to deal with intercepted trasmit mbufs which
151  *   require offloadings like TSO, UFO, checksumming offloadings, etc. It is
152  *   responsibility of the user to disable those offloadings (e.g. using
153  *   ifconfig on FreeBSD or ethtool -K on Linux) for an interface that is being
154  *   used in netmap mode. If the offloadings are not disabled, GSO and/or
155  *   unchecksummed packets may be dropped immediately or end up in the host RX
156  *   rings, and will be dropped as soon as the packet reaches another netmap
157  *   adapter.
158  */
159 
160 /*
161  * struct netmap_slot is a buffer descriptor
162  */
163 struct netmap_slot {
164 	uint32_t buf_idx;	/* buffer index */
165 	uint16_t len;		/* length for this slot */
166 	uint16_t flags;		/* buf changed, etc. */
167 	uint64_t ptr;		/* pointer for indirect buffers */
168 };
169 
170 /*
171  * The following flags control how the slot is used
172  */
173 
174 #define	NS_BUF_CHANGED	0x0001	/* buf_idx changed */
175 	/*
176 	 * must be set whenever buf_idx is changed (as it might be
177 	 * necessary to recompute the physical address and mapping)
178 	 *
179 	 * It is also set by the kernel whenever the buf_idx is
180 	 * changed internally (e.g., by pipes). Applications may
181 	 * use this information to know when they can reuse the
182 	 * contents of previously prepared buffers.
183 	 */
184 
185 #define	NS_REPORT	0x0002	/* ask the hardware to report results */
186 	/*
187 	 * Request notification when slot is used by the hardware.
188 	 * Normally transmit completions are handled lazily and
189 	 * may be unreported. This flag lets us know when a slot
190 	 * has been sent (e.g. to terminate the sender).
191 	 */
192 
193 #define	NS_FORWARD	0x0004	/* pass packet 'forward' */
194 	/*
195 	 * (Only for physical ports, rx rings with NR_FORWARD set).
196 	 * Slot released to the kernel (i.e. before ring->head) with
197 	 * this flag set are passed to the peer ring (host/NIC),
198 	 * thus restoring the host-NIC connection for these slots.
199 	 * This supports efficient traffic monitoring or firewalling.
200 	 */
201 
202 #define	NS_NO_LEARN	0x0008	/* disable bridge learning */
203  	/*
204 	 * On a VALE switch, do not 'learn' the source port for
205  	 * this buffer.
206 	 */
207 
208 #define	NS_INDIRECT	0x0010	/* userspace buffer */
209  	/*
210 	 * (VALE tx rings only) data is in a userspace buffer,
211 	 * whose address is in the 'ptr' field in the slot.
212 	 */
213 
214 #define	NS_MOREFRAG	0x0020	/* packet has more fragments */
215  	/*
216 	 * (VALE ports, ptnetmap ports and some NIC ports, e.g.
217          * ixgbe and i40e on Linux)
218 	 * Set on all but the last slot of a multi-segment packet.
219 	 * The 'len' field refers to the individual fragment.
220 	 */
221 
222 #define	NS_PORT_SHIFT	8
223 #define	NS_PORT_MASK	(0xff << NS_PORT_SHIFT)
224 	/*
225  	 * The high 8 bits of the flag, if not zero, indicate the
226 	 * destination port for the VALE switch, overriding
227  	 * the lookup table.
228  	 */
229 
230 #define	NS_RFRAGS(_slot)	( ((_slot)->flags >> 8) & 0xff)
231 	/*
232 	 * (VALE rx rings only) the high 8 bits
233 	 *  are the number of fragments.
234 	 */
235 
236 #define NETMAP_MAX_FRAGS	64	/* max number of fragments */
237 
238 
239 /*
240  * struct netmap_ring
241  *
242  * Netmap representation of a TX or RX ring (also known as "queue").
243  * This is a queue implemented as a fixed-size circular array.
244  * At the software level the important fields are: head, cur, tail.
245  *
246  * In TX rings:
247  *
248  *	head	first slot available for transmission.
249  *	cur	wakeup point. select() and poll() will unblock
250  *		when 'tail' moves past 'cur'
251  *	tail	(readonly) first slot reserved to the kernel
252  *
253  *	[head .. tail-1] can be used for new packets to send;
254  *	'head' and 'cur' must be incremented as slots are filled
255  *	    with new packets to be sent;
256  *	'cur' can be moved further ahead if we need more space
257  *	for new transmissions. XXX todo (2014-03-12)
258  *
259  * In RX rings:
260  *
261  *	head	first valid received packet
262  *	cur	wakeup point. select() and poll() will unblock
263  *		when 'tail' moves past 'cur'
264  *	tail	(readonly) first slot reserved to the kernel
265  *
266  *	[head .. tail-1] contain received packets;
267  *	'head' and 'cur' must be incremented as slots are consumed
268  *		and can be returned to the kernel;
269  *	'cur' can be moved further ahead if we want to wait for
270  *		new packets without returning the previous ones.
271  *
272  * DATA OWNERSHIP/LOCKING:
273  *	The netmap_ring, and all slots and buffers in the range
274  *	[head .. tail-1] are owned by the user program;
275  *	the kernel only accesses them during a netmap system call
276  *	and in the user thread context.
277  *
278  *	Other slots and buffers are reserved for use by the kernel
279  */
280 struct netmap_ring {
281 	/*
282 	 * buf_ofs is meant to be used through macros.
283 	 * It contains the offset of the buffer region from this
284 	 * descriptor.
285 	 */
286 	const int64_t	buf_ofs;
287 	const uint32_t	num_slots;	/* number of slots in the ring. */
288 	const uint32_t	nr_buf_size;
289 	const uint16_t	ringid;
290 	const uint16_t	dir;		/* 0: tx, 1: rx */
291 
292 	uint32_t        head;		/* (u) first user slot */
293 	uint32_t        cur;		/* (u) wakeup point */
294 	uint32_t	tail;		/* (k) first kernel slot */
295 
296 	uint32_t	flags;
297 
298 	struct timeval	ts;		/* (k) time of last *sync() */
299 
300 	/* opaque room for a mutex or similar object */
301 #if !defined(_WIN32) || defined(__CYGWIN__)
302 	uint8_t	__attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128];
303 #else
304 	uint8_t	__declspec(align(NM_CACHE_ALIGN)) sem[128];
305 #endif
306 
307 	/* the slots follow. This struct has variable size */
308 	struct netmap_slot slot[0];	/* array of slots. */
309 };
310 
311 
312 /*
313  * RING FLAGS
314  */
315 #define	NR_TIMESTAMP	0x0002		/* set timestamp on *sync() */
316 	/*
317 	 * updates the 'ts' field on each netmap syscall. This saves
318 	 * saves a separate gettimeofday(), and is not much worse than
319 	 * software timestamps generated in the interrupt handler.
320 	 */
321 
322 #define	NR_FORWARD	0x0004		/* enable NS_FORWARD for ring */
323  	/*
324 	 * Enables the NS_FORWARD slot flag for the ring.
325 	 */
326 
327 /*
328  * Helper functions for kernel and userspace
329  */
330 
331 /*
332  * Check if space is available in the ring. We use ring->head, which
333  * points to the next netmap slot to be published to netmap. It is
334  * possible that the applications moves ring->cur ahead of ring->tail
335  * (e.g., by setting ring->cur <== ring->tail), if it wants more slots
336  * than the ones currently available, and it wants to be notified when
337  * more arrive. See netmap(4) for more details and examples.
338  */
339 static inline int
340 nm_ring_empty(struct netmap_ring *ring)
341 {
342 	return (ring->head == ring->tail);
343 }
344 
345 /*
346  * Netmap representation of an interface and its queue(s).
347  * This is initialized by the kernel when binding a file
348  * descriptor to a port, and should be considered as readonly
349  * by user programs. The kernel never uses it.
350  *
351  * There is one netmap_if for each file descriptor on which we want
352  * to select/poll.
353  * select/poll operates on one or all pairs depending on the value of
354  * nmr_queueid passed on the ioctl.
355  */
356 struct netmap_if {
357 	char		ni_name[IFNAMSIZ]; /* name of the interface. */
358 	const uint32_t	ni_version;	/* API version, currently unused */
359 	const uint32_t	ni_flags;	/* properties */
360 #define	NI_PRIV_MEM	0x1		/* private memory region */
361 
362 	/*
363 	 * The number of packet rings available in netmap mode.
364 	 * Physical NICs can have different numbers of tx and rx rings.
365 	 * Physical NICs also have at least a 'host' rings pair.
366 	 * Additionally, clients can request additional ring pairs to
367 	 * be used for internal communication.
368 	 */
369 	const uint32_t	ni_tx_rings;	/* number of HW tx rings */
370 	const uint32_t	ni_rx_rings;	/* number of HW rx rings */
371 
372 	uint32_t	ni_bufs_head;	/* head index for extra bufs */
373 	const uint32_t	ni_host_tx_rings; /* number of SW tx rings */
374 	const uint32_t	ni_host_rx_rings; /* number of SW rx rings */
375 	uint32_t	ni_spare1[3];
376 	/*
377 	 * The following array contains the offset of each netmap ring
378 	 * from this structure, in the following order:
379 	 *     - NIC tx rings (ni_tx_rings);
380 	 *     - host tx rings (ni_host_tx_rings);
381 	 *     - NIC rx rings (ni_rx_rings);
382 	 *     - host rx ring (ni_host_rx_rings);
383 	 *
384 	 * The area is filled up by the kernel on NETMAP_REQ_REGISTER,
385 	 * and then only read by userspace code.
386 	 */
387 	const ssize_t	ring_ofs[0];
388 };
389 
390 /* Legacy interface to interact with a netmap control device.
391  * Included for backward compatibility. The user should not include this
392  * file directly. */
393 #include "netmap_legacy.h"
394 
395 /*
396  * New API to control netmap control devices. New applications should only use
397  * nmreq_xyz structs with the NIOCCTRL ioctl() command.
398  *
399  * NIOCCTRL takes a nmreq_header struct, which contains the required
400  * API version, the name of a netmap port, a command type, and pointers
401  * to request body and options.
402  *
403  *	nr_name	(in)
404  *		The name of the port (em0, valeXXX:YYY, eth0{pn1 etc.)
405  *
406  *	nr_version (in/out)
407  *		Must match NETMAP_API as used in the kernel, error otherwise.
408  *		Always returns the desired value on output.
409  *
410  *	nr_reqtype (in)
411  *		One of the NETMAP_REQ_* command types below
412  *
413  *	nr_body (in)
414  *		Pointer to a command-specific struct, described by one
415  *		of the struct nmreq_xyz below.
416  *
417  *	nr_options (in)
418  *		Command specific options, if any.
419  *
420  * A NETMAP_REQ_REGISTER command activates netmap mode on the netmap
421  * port (e.g. physical interface) specified by nmreq_header.nr_name.
422  * The request body (struct nmreq_register) has several arguments to
423  * specify how the port is to be registered.
424  *
425  *	nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings,
426  *	nr_host_tx_rings, nr_host_rx_rings (in/out)
427  *		On input, non-zero values may be used to reconfigure the port
428  *		according to the requested values, but this is not guaranteed.
429  *		On output the actual values in use are reported.
430  *
431  *	nr_mode (in)
432  *		Indicate what set of rings must be bound to the netmap
433  *		device (e.g. all NIC rings, host rings only, NIC and
434  *		host rings, ...). Values are in NR_REG_*.
435  *
436  *	nr_ringid (in)
437  *		If nr_mode == NR_REG_ONE_NIC (only a single couple of TX/RX
438  *		rings), indicate which NIC TX and/or RX ring is to be bound
439  *		(0..nr_*x_rings-1).
440  *
441  *	nr_flags (in)
442  *		Indicate special options for how to open the port.
443  *
444  *		NR_NO_TX_POLL can be OR-ed to make select()/poll() push
445  *			packets on tx rings only if POLLOUT is set.
446  *			The default is to push any pending packet.
447  *
448  *		NR_DO_RX_POLL can be OR-ed to make select()/poll() release
449  *			packets on rx rings also when POLLIN is NOT set.
450  *			The default is to touch the rx ring only with POLLIN.
451  *			Note that this is the opposite of TX because it
452  *			reflects the common usage.
453  *
454  *		Other options are NR_MONITOR_TX, NR_MONITOR_RX, NR_ZCOPY_MON,
455  *		NR_EXCLUSIVE, NR_RX_RINGS_ONLY, NR_TX_RINGS_ONLY and
456  *		NR_ACCEPT_VNET_HDR.
457  *
458  *	nr_mem_id (in/out)
459  *		The identity of the memory region used.
460  *		On input, 0 means the system decides autonomously,
461  *		other values may try to select a specific region.
462  *		On return the actual value is reported.
463  *		Region '1' is the global allocator, normally shared
464  *		by all interfaces. Other values are private regions.
465  *		If two ports the same region zero-copy is possible.
466  *
467  *	nr_extra_bufs (in/out)
468  *		Number of extra buffers to be allocated.
469  *
470  * The other NETMAP_REQ_* commands are described below.
471  *
472  */
473 
474 /* maximum size of a request, including all options */
475 #define NETMAP_REQ_MAXSIZE	4096
476 
477 /* Header common to all request options. */
478 struct nmreq_option {
479 	/* Pointer ot the next option. */
480 	uint64_t		nro_next;
481 	/* Option type. */
482 	uint32_t		nro_reqtype;
483 	/* (out) status of the option:
484 	 * 0: recognized and processed
485 	 * !=0: errno value
486 	 */
487 	uint32_t		nro_status;
488 	/* Option size, used only for options that can have variable size
489 	 * (e.g. because they contain arrays). For fixed-size options this
490 	 * field should be set to zero. */
491 	uint64_t		nro_size;
492 };
493 
494 /* Header common to all requests. Do not reorder these fields, as we need
495  * the second one (nr_reqtype) to know how much to copy from/to userspace. */
496 struct nmreq_header {
497 	uint16_t		nr_version;	/* API version */
498 	uint16_t		nr_reqtype;	/* nmreq type (NETMAP_REQ_*) */
499 	uint32_t		nr_reserved;	/* must be zero */
500 #define NETMAP_REQ_IFNAMSIZ	64
501 	char			nr_name[NETMAP_REQ_IFNAMSIZ]; /* port name */
502 	uint64_t		nr_options;	/* command-specific options */
503 	uint64_t		nr_body;	/* ptr to nmreq_xyz struct */
504 };
505 
506 enum {
507 	/* Register a netmap port with the device. */
508 	NETMAP_REQ_REGISTER = 1,
509 	/* Get information from a netmap port. */
510 	NETMAP_REQ_PORT_INFO_GET,
511 	/* Attach a netmap port to a VALE switch. */
512 	NETMAP_REQ_VALE_ATTACH,
513 	/* Detach a netmap port from a VALE switch. */
514 	NETMAP_REQ_VALE_DETACH,
515 	/* List the ports attached to a VALE switch. */
516 	NETMAP_REQ_VALE_LIST,
517 	/* Set the port header length (was virtio-net header length). */
518 	NETMAP_REQ_PORT_HDR_SET,
519 	/* Get the port header length (was virtio-net header length). */
520 	NETMAP_REQ_PORT_HDR_GET,
521 	/* Create a new persistent VALE port. */
522 	NETMAP_REQ_VALE_NEWIF,
523 	/* Delete a persistent VALE port. */
524 	NETMAP_REQ_VALE_DELIF,
525 	/* Enable polling kernel thread(s) on an attached VALE port. */
526 	NETMAP_REQ_VALE_POLLING_ENABLE,
527 	/* Disable polling kernel thread(s) on an attached VALE port. */
528 	NETMAP_REQ_VALE_POLLING_DISABLE,
529 	/* Get info about the pools of a memory allocator. */
530 	NETMAP_REQ_POOLS_INFO_GET,
531 	/* Start an in-kernel loop that syncs the rings periodically or
532 	 * on notifications. The loop runs in the context of the ioctl
533 	 * syscall, and only stops on NETMAP_REQ_SYNC_KLOOP_STOP. */
534 	NETMAP_REQ_SYNC_KLOOP_START,
535 	/* Stops the thread executing the in-kernel loop. The thread
536 	 * returns from the ioctl syscall. */
537 	NETMAP_REQ_SYNC_KLOOP_STOP,
538 	/* Enable CSB mode on a registered netmap control device. */
539 	NETMAP_REQ_CSB_ENABLE,
540 };
541 
542 enum {
543 	/* On NETMAP_REQ_REGISTER, ask netmap to use memory allocated
544 	 * from user-space allocated memory pools (e.g. hugepages).
545 	 */
546 	NETMAP_REQ_OPT_EXTMEM = 1,
547 
548 	/* ON NETMAP_REQ_SYNC_KLOOP_START, ask netmap to use eventfd-based
549 	 * notifications to synchronize the kernel loop with the application.
550 	 */
551 	NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS,
552 
553 	/* On NETMAP_REQ_REGISTER, ask netmap to work in CSB mode, where
554 	 * head, cur and tail pointers are not exchanged through the
555 	 * struct netmap_ring header, but rather using an user-provided
556 	 * memory area (see struct nm_csb_atok and struct nm_csb_ktoa).
557 	 */
558 	NETMAP_REQ_OPT_CSB,
559 
560 	/* An extension to NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS, which specifies
561 	 * if the TX and/or RX rings are synced in the context of the VM exit.
562 	 * This requires the 'ioeventfd' fields to be valid (cannot be < 0).
563 	 */
564 	NETMAP_REQ_OPT_SYNC_KLOOP_MODE,
565 };
566 
567 /*
568  * nr_reqtype: NETMAP_REQ_REGISTER
569  * Bind (register) a netmap port to this control device.
570  */
571 struct nmreq_register {
572 	uint64_t	nr_offset;	/* nifp offset in the shared region */
573 	uint64_t	nr_memsize;	/* size of the shared region */
574 	uint32_t	nr_tx_slots;	/* slots in tx rings */
575 	uint32_t	nr_rx_slots;	/* slots in rx rings */
576 	uint16_t	nr_tx_rings;	/* number of tx rings */
577 	uint16_t	nr_rx_rings;	/* number of rx rings */
578 	uint16_t	nr_host_tx_rings; /* number of host tx rings */
579 	uint16_t	nr_host_rx_rings; /* number of host rx rings */
580 
581 	uint16_t	nr_mem_id;	/* id of the memory allocator */
582 	uint16_t	nr_ringid;	/* ring(s) we care about */
583 	uint32_t	nr_mode;	/* specify NR_REG_* modes */
584 	uint32_t	nr_extra_bufs;	/* number of requested extra buffers */
585 
586 	uint64_t	nr_flags;	/* additional flags (see below) */
587 /* monitors use nr_ringid and nr_mode to select the rings to monitor */
588 #define NR_MONITOR_TX	0x100
589 #define NR_MONITOR_RX	0x200
590 #define NR_ZCOPY_MON	0x400
591 /* request exclusive access to the selected rings */
592 #define NR_EXCLUSIVE	0x800
593 /* 0x1000 unused */
594 #define NR_RX_RINGS_ONLY	0x2000
595 #define NR_TX_RINGS_ONLY	0x4000
596 /* Applications set this flag if they are able to deal with virtio-net headers,
597  * that is send/receive frames that start with a virtio-net header.
598  * If not set, NETMAP_REQ_REGISTER will fail with netmap ports that require
599  * applications to use those headers. If the flag is set, the application can
600  * use the NETMAP_VNET_HDR_GET command to figure out the header length. */
601 #define NR_ACCEPT_VNET_HDR	0x8000
602 /* The following two have the same meaning of NETMAP_NO_TX_POLL and
603  * NETMAP_DO_RX_POLL. */
604 #define NR_DO_RX_POLL		0x10000
605 #define NR_NO_TX_POLL		0x20000
606 };
607 
608 /* Valid values for nmreq_register.nr_mode (see above). */
609 enum {	NR_REG_DEFAULT	= 0,	/* backward compat, should not be used. */
610 	NR_REG_ALL_NIC	= 1,
611 	NR_REG_SW	= 2,
612 	NR_REG_NIC_SW	= 3,
613 	NR_REG_ONE_NIC	= 4,
614 	NR_REG_PIPE_MASTER = 5, /* deprecated, use "x{y" port name syntax */
615 	NR_REG_PIPE_SLAVE = 6,  /* deprecated, use "x}y" port name syntax */
616 	NR_REG_NULL     = 7,
617 	NR_REG_ONE_SW	= 8,
618 };
619 
620 /* A single ioctl number is shared by all the new API command.
621  * Demultiplexing is done using the hdr.nr_reqtype field.
622  * FreeBSD uses the size value embedded in the _IOWR to determine
623  * how much to copy in/out, so we define the ioctl() command
624  * specifying only nmreq_header, and copyin/copyout the rest. */
625 #define NIOCCTRL	_IOWR('i', 151, struct nmreq_header)
626 
627 /* The ioctl commands to sync TX/RX netmap rings.
628  * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
629  *	whose identity is set in NETMAP_REQ_REGISTER through nr_ringid.
630  *	These are non blocking and take no argument. */
631 #define NIOCTXSYNC	_IO('i', 148) /* sync tx queues */
632 #define NIOCRXSYNC	_IO('i', 149) /* sync rx queues */
633 
634 /*
635  * nr_reqtype: NETMAP_REQ_PORT_INFO_GET
636  * Get information about a netmap port, including number of rings.
637  * slots per ring, id of the memory allocator, etc. The netmap
638  * control device used for this operation does not need to be bound
639  * to a netmap port.
640  */
641 struct nmreq_port_info_get {
642 	uint64_t	nr_memsize;	/* size of the shared region */
643 	uint32_t	nr_tx_slots;	/* slots in tx rings */
644 	uint32_t	nr_rx_slots;	/* slots in rx rings */
645 	uint16_t	nr_tx_rings;	/* number of tx rings */
646 	uint16_t	nr_rx_rings;	/* number of rx rings */
647 	uint16_t	nr_host_tx_rings; /* number of host tx rings */
648 	uint16_t	nr_host_rx_rings; /* number of host rx rings */
649 	uint16_t	nr_mem_id;	/* memory allocator id (in/out) */
650 	uint16_t	pad[3];
651 };
652 
653 #define	NM_BDG_NAME		"vale"	/* prefix for bridge port name */
654 
655 /*
656  * nr_reqtype: NETMAP_REQ_VALE_ATTACH
657  * Attach a netmap port to a VALE switch. Both the name of the netmap
658  * port and the VALE switch are specified through the nr_name argument.
659  * The attach operation could need to register a port, so at least
660  * the same arguments are available.
661  * port_index will contain the index where the port has been attached.
662  */
663 struct nmreq_vale_attach {
664 	struct nmreq_register reg;
665 	uint32_t port_index;
666 	uint32_t pad1;
667 };
668 
669 /*
670  * nr_reqtype: NETMAP_REQ_VALE_DETACH
671  * Detach a netmap port from a VALE switch. Both the name of the netmap
672  * port and the VALE switch are specified through the nr_name argument.
673  * port_index will contain the index where the port was attached.
674  */
675 struct nmreq_vale_detach {
676 	uint32_t port_index;
677 	uint32_t pad1;
678 };
679 
680 /*
681  * nr_reqtype: NETMAP_REQ_VALE_LIST
682  * List the ports of a VALE switch.
683  */
684 struct nmreq_vale_list {
685 	/* Name of the VALE port (valeXXX:YYY) or empty. */
686 	uint16_t	nr_bridge_idx;
687 	uint16_t	pad1;
688 	uint32_t	nr_port_idx;
689 };
690 
691 /*
692  * nr_reqtype: NETMAP_REQ_PORT_HDR_SET or NETMAP_REQ_PORT_HDR_GET
693  * Set or get the port header length of the port identified by hdr.nr_name.
694  * The control device does not need to be bound to a netmap port.
695  */
696 struct nmreq_port_hdr {
697 	uint32_t	nr_hdr_len;
698 	uint32_t	pad1;
699 };
700 
701 /*
702  * nr_reqtype: NETMAP_REQ_VALE_NEWIF
703  * Create a new persistent VALE port.
704  */
705 struct nmreq_vale_newif {
706 	uint32_t	nr_tx_slots;	/* slots in tx rings */
707 	uint32_t	nr_rx_slots;	/* slots in rx rings */
708 	uint16_t	nr_tx_rings;	/* number of tx rings */
709 	uint16_t	nr_rx_rings;	/* number of rx rings */
710 	uint16_t	nr_mem_id;	/* id of the memory allocator */
711 	uint16_t	pad1;
712 };
713 
714 /*
715  * nr_reqtype: NETMAP_REQ_VALE_POLLING_ENABLE or NETMAP_REQ_VALE_POLLING_DISABLE
716  * Enable or disable polling kthreads on a VALE port.
717  */
718 struct nmreq_vale_polling {
719 	uint32_t	nr_mode;
720 #define NETMAP_POLLING_MODE_SINGLE_CPU 1
721 #define NETMAP_POLLING_MODE_MULTI_CPU 2
722 	uint32_t	nr_first_cpu_id;
723 	uint32_t	nr_num_polling_cpus;
724 	uint32_t	pad1;
725 };
726 
727 /*
728  * nr_reqtype: NETMAP_REQ_POOLS_INFO_GET
729  * Get info about the pools of the memory allocator of the netmap
730  * port specified by hdr.nr_name and nr_mem_id. The netmap control
731  * device used for this operation does not need to be bound to a netmap
732  * port.
733  */
734 struct nmreq_pools_info {
735 	uint64_t	nr_memsize;
736 	uint16_t	nr_mem_id; /* in/out argument */
737 	uint16_t	pad1[3];
738 	uint64_t	nr_if_pool_offset;
739 	uint32_t	nr_if_pool_objtotal;
740 	uint32_t	nr_if_pool_objsize;
741 	uint64_t	nr_ring_pool_offset;
742 	uint32_t	nr_ring_pool_objtotal;
743 	uint32_t	nr_ring_pool_objsize;
744 	uint64_t	nr_buf_pool_offset;
745 	uint32_t	nr_buf_pool_objtotal;
746 	uint32_t	nr_buf_pool_objsize;
747 };
748 
749 /*
750  * nr_reqtype: NETMAP_REQ_SYNC_KLOOP_START
751  * Start an in-kernel loop that syncs the rings periodically or on
752  * notifications. The loop runs in the context of the ioctl syscall,
753  * and only stops on NETMAP_REQ_SYNC_KLOOP_STOP.
754  * The registered netmap port must be open in CSB mode.
755  */
756 struct nmreq_sync_kloop_start {
757 	/* Sleeping is the default synchronization method for the kloop.
758 	 * The 'sleep_us' field specifies how many microsconds to sleep for
759 	 * when there is no work to do, before doing another kloop iteration.
760 	 */
761 	uint32_t	sleep_us;
762 	uint32_t	pad1;
763 };
764 
765 /* A CSB entry for the application --> kernel direction. */
766 struct nm_csb_atok {
767 	uint32_t head;		  /* AW+ KR+ the head of the appl netmap_ring */
768 	uint32_t cur;		  /* AW+ KR+ the cur of the appl netmap_ring */
769 	uint32_t appl_need_kick;  /* AW+ KR+ kern --> appl notification enable */
770 	uint32_t sync_flags;	  /* AW+ KR+ the flags of the appl [tx|rx]sync() */
771 	uint32_t pad[12];	  /* pad to a 64 bytes cacheline */
772 };
773 
774 /* A CSB entry for the application <-- kernel direction. */
775 struct nm_csb_ktoa {
776 	uint32_t hwcur;		  /* AR+ KW+ the hwcur of the kern netmap_kring */
777 	uint32_t hwtail;	  /* AR+ KW+ the hwtail of the kern netmap_kring */
778 	uint32_t kern_need_kick;  /* AR+ KW+ appl-->kern notification enable */
779 	uint32_t pad[13];
780 };
781 
782 #ifdef __linux__
783 
784 #ifdef __KERNEL__
785 #define nm_stst_barrier smp_wmb
786 #define nm_ldld_barrier smp_rmb
787 #define nm_stld_barrier smp_mb
788 #else  /* !__KERNEL__ */
789 static inline void nm_stst_barrier(void)
790 {
791 	/* A memory barrier with release semantic has the combined
792 	 * effect of a store-store barrier and a load-store barrier,
793 	 * which is fine for us. */
794 	__atomic_thread_fence(__ATOMIC_RELEASE);
795 }
796 static inline void nm_ldld_barrier(void)
797 {
798 	/* A memory barrier with acquire semantic has the combined
799 	 * effect of a load-load barrier and a store-load barrier,
800 	 * which is fine for us. */
801 	__atomic_thread_fence(__ATOMIC_ACQUIRE);
802 }
803 #endif /* !__KERNEL__ */
804 
805 #elif defined(__FreeBSD__)
806 
807 #ifdef _KERNEL
808 #define nm_stst_barrier	atomic_thread_fence_rel
809 #define nm_ldld_barrier	atomic_thread_fence_acq
810 #define nm_stld_barrier	atomic_thread_fence_seq_cst
811 #else  /* !_KERNEL */
812 #include <stdatomic.h>
813 static inline void nm_stst_barrier(void)
814 {
815 	atomic_thread_fence(memory_order_release);
816 }
817 static inline void nm_ldld_barrier(void)
818 {
819 	atomic_thread_fence(memory_order_acquire);
820 }
821 #endif /* !_KERNEL */
822 
823 #else  /* !__linux__ && !__FreeBSD__ */
824 #error "OS not supported"
825 #endif /* !__linux__ && !__FreeBSD__ */
826 
827 /* Application side of sync-kloop: Write ring pointers (cur, head) to the CSB.
828  * This routine is coupled with sync_kloop_kernel_read(). */
829 static inline void
830 nm_sync_kloop_appl_write(struct nm_csb_atok *atok, uint32_t cur,
831 			 uint32_t head)
832 {
833 	/* Issue a first store-store barrier to make sure writes to the
834 	 * netmap ring do not overcome updates on atok->cur and atok->head. */
835 	nm_stst_barrier();
836 
837 	/*
838 	 * We need to write cur and head to the CSB but we cannot do it atomically.
839 	 * There is no way we can prevent the host from reading the updated value
840 	 * of one of the two and the old value of the other. However, if we make
841 	 * sure that the host never reads a value of head more recent than the
842 	 * value of cur we are safe. We can allow the host to read a value of cur
843 	 * more recent than the value of head, since in the netmap ring cur can be
844 	 * ahead of head and cur cannot wrap around head because it must be behind
845 	 * tail. Inverting the order of writes below could instead result into the
846 	 * host to think head went ahead of cur, which would cause the sync
847 	 * prologue to fail.
848 	 *
849 	 * The following memory barrier scheme is used to make this happen:
850 	 *
851 	 *          Guest                Host
852 	 *
853 	 *          STORE(cur)           LOAD(head)
854 	 *          wmb() <----------->  rmb()
855 	 *          STORE(head)          LOAD(cur)
856 	 *
857 	 */
858 	atok->cur = cur;
859 	nm_stst_barrier();
860 	atok->head = head;
861 }
862 
863 /* Application side of sync-kloop: Read kring pointers (hwcur, hwtail) from
864  * the CSB. This routine is coupled with sync_kloop_kernel_write(). */
865 static inline void
866 nm_sync_kloop_appl_read(struct nm_csb_ktoa *ktoa, uint32_t *hwtail,
867 			uint32_t *hwcur)
868 {
869 	/*
870 	 * We place a memory barrier to make sure that the update of hwtail never
871 	 * overtakes the update of hwcur.
872 	 * (see explanation in sync_kloop_kernel_write).
873 	 */
874 	*hwtail = ktoa->hwtail;
875 	nm_ldld_barrier();
876 	*hwcur = ktoa->hwcur;
877 
878 	/* Make sure that loads from ktoa->hwtail and ktoa->hwcur are not delayed
879 	 * after the loads from the netmap ring. */
880 	nm_ldld_barrier();
881 }
882 
883 /*
884  * data for NETMAP_REQ_OPT_* options
885  */
886 
887 struct nmreq_opt_sync_kloop_eventfds {
888 	struct nmreq_option	nro_opt;	/* common header */
889 	/* An array of N entries for bidirectional notifications between
890 	 * the kernel loop and the application. The number of entries and
891 	 * their order must agree with the CSB arrays passed in the
892 	 * NETMAP_REQ_OPT_CSB option. Each entry contains a file descriptor
893 	 * backed by an eventfd.
894 	 *
895 	 * If any of the 'ioeventfd' entries is < 0, the event loop uses
896 	 * the sleeping synchronization strategy (according to sleep_us),
897 	 * and keeps kern_need_kick always disabled.
898 	 * Each 'irqfd' can be < 0, and in that case the corresponding queue
899 	 * is never notified.
900 	 */
901 	struct {
902 		/* Notifier for the application --> kernel loop direction. */
903 		int32_t ioeventfd;
904 		/* Notifier for the kernel loop --> application direction. */
905 		int32_t irqfd;
906 	} eventfds[0];
907 };
908 
909 struct nmreq_opt_sync_kloop_mode {
910 	struct nmreq_option	nro_opt;	/* common header */
911 #define NM_OPT_SYNC_KLOOP_DIRECT_TX (1 << 0)
912 #define NM_OPT_SYNC_KLOOP_DIRECT_RX (1 << 1)
913 	uint32_t mode;
914 };
915 
916 struct nmreq_opt_extmem {
917 	struct nmreq_option	nro_opt;	/* common header */
918 	uint64_t		nro_usrptr;	/* (in) ptr to usr memory */
919 	struct nmreq_pools_info	nro_info;	/* (in/out) */
920 };
921 
922 struct nmreq_opt_csb {
923 	struct nmreq_option	nro_opt;
924 
925 	/* Array of CSB entries for application --> kernel communication
926 	 * (N entries). */
927 	uint64_t		csb_atok;
928 
929 	/* Array of CSB entries for kernel --> application communication
930 	 * (N entries). */
931 	uint64_t		csb_ktoa;
932 };
933 
934 #endif /* _NET_NETMAP_H_ */
935