xref: /freebsd/sys/net/netmap_user.h (revision 6829dae12bb055451fa467da4589c43bd03b1e64)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2011-2016 Universita` di Pisa
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  *   1. Redistributions of source code must retain the above copyright
12  *      notice, this list of conditions and the following disclaimer.
13  *   2. Redistributions in binary form must reproduce the above copyright
14  *      notice, this list of conditions and the following disclaimer in the
15  *      documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * $FreeBSD$
32  *
33  * Functions and macros to manipulate netmap structures and packets
34  * in userspace. See netmap(4) for more information.
35  *
36  * The address of the struct netmap_if, say nifp, is computed from the
37  * value returned from ioctl(.., NIOCREG, ...) and the mmap region:
38  *	ioctl(fd, NIOCREG, &req);
39  *	mem = mmap(0, ... );
40  *	nifp = NETMAP_IF(mem, req.nr_nifp);
41  *		(so simple, we could just do it manually)
42  *
43  * From there:
44  *	struct netmap_ring *NETMAP_TXRING(nifp, index)
45  *	struct netmap_ring *NETMAP_RXRING(nifp, index)
46  *		we can access ring->cur, ring->head, ring->tail, etc.
47  *
48  *	ring->slot[i] gives us the i-th slot (we can access
49  *		directly len, flags, buf_idx)
50  *
51  *	char *buf = NETMAP_BUF(ring, x) returns a pointer to
52  *		the buffer numbered x
53  *
54  * All ring indexes (head, cur, tail) should always move forward.
55  * To compute the next index in a circular ring you can use
56  *	i = nm_ring_next(ring, i);
57  *
58  * To ease porting apps from pcap to netmap we supply a few fuctions
59  * that can be called to open, close, read and write on netmap in a way
60  * similar to libpcap. Note that the read/write function depend on
61  * an ioctl()/select()/poll() being issued to refill rings or push
62  * packets out.
63  *
64  * In order to use these, include #define NETMAP_WITH_LIBS
65  * in the source file that invokes these functions.
66  */
67 
68 #ifndef _NET_NETMAP_USER_H_
69 #define _NET_NETMAP_USER_H_
70 
71 #define NETMAP_DEVICE_NAME "/dev/netmap"
72 
73 #ifdef __CYGWIN__
74 /*
75  * we can compile userspace apps with either cygwin or msvc,
76  * and we use _WIN32 to identify windows specific code
77  */
78 #ifndef _WIN32
79 #define _WIN32
80 #endif	/* _WIN32 */
81 
82 #endif	/* __CYGWIN__ */
83 
84 #ifdef _WIN32
85 #undef NETMAP_DEVICE_NAME
86 #define NETMAP_DEVICE_NAME "/proc/sys/DosDevices/Global/netmap"
87 #include <windows.h>
88 #include <WinDef.h>
89 #include <sys/cygwin.h>
90 #endif /* _WIN32 */
91 
92 #include <stdint.h>
93 #include <sys/socket.h>		/* apple needs sockaddr */
94 #include <net/if.h>		/* IFNAMSIZ */
95 #include <ctype.h>
96 
97 #ifndef likely
98 #define likely(x)	__builtin_expect(!!(x), 1)
99 #define unlikely(x)	__builtin_expect(!!(x), 0)
100 #endif /* likely and unlikely */
101 
102 #include <net/netmap.h>
103 
104 /* helper macro */
105 #define _NETMAP_OFFSET(type, ptr, offset) \
106 	((type)(void *)((char *)(ptr) + (offset)))
107 
108 #define NETMAP_IF(_base, _ofs)	_NETMAP_OFFSET(struct netmap_if *, _base, _ofs)
109 
110 #define NETMAP_TXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *, \
111 	nifp, (nifp)->ring_ofs[index] )
112 
113 #define NETMAP_RXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *,	\
114 	nifp, (nifp)->ring_ofs[index + (nifp)->ni_tx_rings + 1] )
115 
116 #define NETMAP_BUF(ring, index)				\
117 	((char *)(ring) + (ring)->buf_ofs + ((index)*(ring)->nr_buf_size))
118 
119 #define NETMAP_BUF_IDX(ring, buf)			\
120 	( ((char *)(buf) - ((char *)(ring) + (ring)->buf_ofs) ) / \
121 		(ring)->nr_buf_size )
122 
123 
124 static inline uint32_t
125 nm_ring_next(struct netmap_ring *r, uint32_t i)
126 {
127 	return ( unlikely(i + 1 == r->num_slots) ? 0 : i + 1);
128 }
129 
130 
131 /*
132  * Return 1 if we have pending transmissions in the tx ring.
133  * When everything is complete ring->head = ring->tail + 1 (modulo ring size)
134  */
135 static inline int
136 nm_tx_pending(struct netmap_ring *r)
137 {
138 	return nm_ring_next(r, r->tail) != r->head;
139 }
140 
141 /* Compute the number of slots available in the netmap ring. We use
142  * ring->head as explained in the comment above nm_ring_empty(). */
143 static inline uint32_t
144 nm_ring_space(struct netmap_ring *ring)
145 {
146         int ret = ring->tail - ring->head;
147         if (ret < 0)
148                 ret += ring->num_slots;
149         return ret;
150 }
151 
152 
153 #ifdef NETMAP_WITH_LIBS
154 /*
155  * Support for simple I/O libraries.
156  * Include other system headers required for compiling this.
157  */
158 
159 #ifndef HAVE_NETMAP_WITH_LIBS
160 #define HAVE_NETMAP_WITH_LIBS
161 
162 #include <stdio.h>
163 #include <sys/time.h>
164 #include <sys/mman.h>
165 #include <string.h>	/* memset */
166 #include <sys/ioctl.h>
167 #include <sys/errno.h>	/* EINVAL */
168 #include <fcntl.h>	/* O_RDWR */
169 #include <unistd.h>	/* close() */
170 #include <signal.h>
171 #include <stdlib.h>
172 
173 #ifndef ND /* debug macros */
174 /* debug support */
175 #define ND(_fmt, ...) do {} while(0)
176 #define D(_fmt, ...)						\
177 	do {							\
178 		struct timeval _t0;				\
179 		gettimeofday(&_t0, NULL);			\
180 		fprintf(stderr, "%03d.%06d %s [%d] " _fmt "\n",	\
181 		    (int)(_t0.tv_sec % 1000), (int)_t0.tv_usec,	\
182 		    __FUNCTION__, __LINE__, ##__VA_ARGS__);	\
183         } while (0)
184 
185 /* Rate limited version of "D", lps indicates how many per second */
186 #define RD(lps, format, ...)                                    \
187     do {                                                        \
188         static int __t0, __cnt;                                 \
189         struct timeval __xxts;                                  \
190         gettimeofday(&__xxts, NULL);                            \
191         if (__t0 != __xxts.tv_sec) {                            \
192             __t0 = __xxts.tv_sec;                               \
193             __cnt = 0;                                          \
194         }                                                       \
195         if (__cnt++ < lps) {                                    \
196             D(format, ##__VA_ARGS__);                           \
197         }                                                       \
198     } while (0)
199 #endif
200 
201 struct nm_pkthdr {	/* first part is the same as pcap_pkthdr */
202 	struct timeval	ts;
203 	uint32_t	caplen;
204 	uint32_t	len;
205 
206 	uint64_t flags;	/* NM_MORE_PKTS etc */
207 #define NM_MORE_PKTS	1
208 	struct nm_desc *d;
209 	struct netmap_slot *slot;
210 	uint8_t *buf;
211 };
212 
213 struct nm_stat {	/* same as pcap_stat	*/
214 	u_int	ps_recv;
215 	u_int	ps_drop;
216 	u_int	ps_ifdrop;
217 #ifdef WIN32 /* XXX or _WIN32 ? */
218 	u_int	bs_capt;
219 #endif /* WIN32 */
220 };
221 
222 #define NM_ERRBUF_SIZE	512
223 
224 struct nm_desc {
225 	struct nm_desc *self; /* point to self if netmap. */
226 	int fd;
227 	void *mem;
228 	uint32_t memsize;
229 	int done_mmap;	/* set if mem is the result of mmap */
230 	struct netmap_if * const nifp;
231 	uint16_t first_tx_ring, last_tx_ring, cur_tx_ring;
232 	uint16_t first_rx_ring, last_rx_ring, cur_rx_ring;
233 	struct nmreq req;	/* also contains the nr_name = ifname */
234 	struct nm_pkthdr hdr;
235 
236 	/*
237 	 * The memory contains netmap_if, rings and then buffers.
238 	 * Given a pointer (e.g. to nm_inject) we can compare with
239 	 * mem/buf_start/buf_end to tell if it is a buffer or
240 	 * some other descriptor in our region.
241 	 * We also store a pointer to some ring as it helps in the
242 	 * translation from buffer indexes to addresses.
243 	 */
244 	struct netmap_ring * const some_ring;
245 	void * const buf_start;
246 	void * const buf_end;
247 	/* parameters from pcap_open_live */
248 	int snaplen;
249 	int promisc;
250 	int to_ms;
251 	char *errbuf;
252 
253 	/* save flags so we can restore them on close */
254 	uint32_t if_flags;
255         uint32_t if_reqcap;
256         uint32_t if_curcap;
257 
258 	struct nm_stat st;
259 	char msg[NM_ERRBUF_SIZE];
260 };
261 
262 /*
263  * when the descriptor is open correctly, d->self == d
264  * Eventually we should also use some magic number.
265  */
266 #define P2NMD(p)		((struct nm_desc *)(p))
267 #define IS_NETMAP_DESC(d)	((d) && P2NMD(d)->self == P2NMD(d))
268 #define NETMAP_FD(d)		(P2NMD(d)->fd)
269 
270 
271 /*
272  * this is a slightly optimized copy routine which rounds
273  * to multiple of 64 bytes and is often faster than dealing
274  * with other odd sizes. We assume there is enough room
275  * in the source and destination buffers.
276  */
277 static inline void
278 nm_pkt_copy(const void *_src, void *_dst, int l)
279 {
280 	const uint64_t *src = (const uint64_t *)_src;
281 	uint64_t *dst = (uint64_t *)_dst;
282 
283 	if (unlikely(l >= 1024 || l % 64)) {
284 		memcpy(dst, src, l);
285 		return;
286 	}
287 	for (; likely(l > 0); l-=64) {
288 		*dst++ = *src++;
289 		*dst++ = *src++;
290 		*dst++ = *src++;
291 		*dst++ = *src++;
292 		*dst++ = *src++;
293 		*dst++ = *src++;
294 		*dst++ = *src++;
295 		*dst++ = *src++;
296 	}
297 }
298 
299 
300 /*
301  * The callback, invoked on each received packet. Same as libpcap
302  */
303 typedef void (*nm_cb_t)(u_char *, const struct nm_pkthdr *, const u_char *d);
304 
305 /*
306  *--- the pcap-like API ---
307  *
308  * nm_open() opens a file descriptor, binds to a port and maps memory.
309  *
310  * ifname	(netmap:foo or vale:foo) is the port name
311  *		a suffix can indicate the follwing:
312  *		^		bind the host (sw) ring pair
313  *		*		bind host and NIC ring pairs
314  *		-NN		bind individual NIC ring pair
315  *		{NN		bind master side of pipe NN
316  *		}NN		bind slave side of pipe NN
317  *		a suffix starting with / and the following flags,
318  *		in any order:
319  *		x		exclusive access
320  *		z		zero copy monitor (both tx and rx)
321  *		t		monitor tx side (copy monitor)
322  *		r		monitor rx side (copy monitor)
323  *		R		bind only RX ring(s)
324  *		T		bind only TX ring(s)
325  *
326  * req		provides the initial values of nmreq before parsing ifname.
327  *		Remember that the ifname parsing will override the ring
328  *		number in nm_ringid, and part of nm_flags;
329  * flags	special functions, normally 0
330  *		indicates which fields of *arg are significant
331  * arg		special functions, normally NULL
332  *		if passed a netmap_desc with mem != NULL,
333  *		use that memory instead of mmap.
334  */
335 
336 static struct nm_desc *nm_open(const char *ifname, const struct nmreq *req,
337 	uint64_t flags, const struct nm_desc *arg);
338 
339 /*
340  * nm_open can import some fields from the parent descriptor.
341  * These flags control which ones.
342  * Also in flags you can specify NETMAP_NO_TX_POLL and NETMAP_DO_RX_POLL,
343  * which set the initial value for these flags.
344  * Note that the 16 low bits of the flags are reserved for data
345  * that may go into the nmreq.
346  */
347 enum {
348 	NM_OPEN_NO_MMAP =	0x040000, /* reuse mmap from parent */
349 	NM_OPEN_IFNAME =	0x080000, /* nr_name, nr_ringid, nr_flags */
350 	NM_OPEN_ARG1 =		0x100000,
351 	NM_OPEN_ARG2 =		0x200000,
352 	NM_OPEN_ARG3 =		0x400000,
353 	NM_OPEN_RING_CFG =	0x800000, /* tx|rx rings|slots */
354 };
355 
356 
357 /*
358  * nm_close()	closes and restores the port to its previous state
359  */
360 
361 static int nm_close(struct nm_desc *);
362 
363 /*
364  * nm_mmap()    do mmap or inherit from parent if the nr_arg2
365  *              (memory block) matches.
366  */
367 
368 static int nm_mmap(struct nm_desc *, const struct nm_desc *);
369 
370 /*
371  * nm_inject() is the same as pcap_inject()
372  * nm_dispatch() is the same as pcap_dispatch()
373  * nm_nextpkt() is the same as pcap_next()
374  */
375 
376 static int nm_inject(struct nm_desc *, const void *, size_t);
377 static int nm_dispatch(struct nm_desc *, int, nm_cb_t, u_char *);
378 static u_char *nm_nextpkt(struct nm_desc *, struct nm_pkthdr *);
379 
380 #ifdef _WIN32
381 
382 intptr_t _get_osfhandle(int); /* defined in io.h in windows */
383 
384 /*
385  * In windows we do not have yet native poll support, so we keep track
386  * of file descriptors associated to netmap ports to emulate poll on
387  * them and fall back on regular poll on other file descriptors.
388  */
389 struct win_netmap_fd_list {
390 	struct win_netmap_fd_list *next;
391 	int win_netmap_fd;
392 	HANDLE win_netmap_handle;
393 };
394 
395 /*
396  * list head containing all the netmap opened fd and their
397  * windows HANDLE counterparts
398  */
399 static struct win_netmap_fd_list *win_netmap_fd_list_head;
400 
401 static void
402 win_insert_fd_record(int fd)
403 {
404 	struct win_netmap_fd_list *curr;
405 
406 	for (curr = win_netmap_fd_list_head; curr; curr = curr->next) {
407 		if (fd == curr->win_netmap_fd) {
408 			return;
409 		}
410 	}
411 	curr = calloc(1, sizeof(*curr));
412 	curr->next = win_netmap_fd_list_head;
413 	curr->win_netmap_fd = fd;
414 	curr->win_netmap_handle = IntToPtr(_get_osfhandle(fd));
415 	win_netmap_fd_list_head = curr;
416 }
417 
418 void
419 win_remove_fd_record(int fd)
420 {
421 	struct win_netmap_fd_list *curr = win_netmap_fd_list_head;
422 	struct win_netmap_fd_list *prev = NULL;
423 	for (; curr ; prev = curr, curr = curr->next) {
424 		if (fd != curr->win_netmap_fd)
425 			continue;
426 		/* found the entry */
427 		if (prev == NULL) { /* we are freeing the first entry */
428 			win_netmap_fd_list_head = curr->next;
429 		} else {
430 			prev->next = curr->next;
431 		}
432 		free(curr);
433 		break;
434 	}
435 }
436 
437 
438 HANDLE
439 win_get_netmap_handle(int fd)
440 {
441 	struct win_netmap_fd_list *curr;
442 
443 	for (curr = win_netmap_fd_list_head; curr; curr = curr->next) {
444 		if (fd == curr->win_netmap_fd) {
445 			return curr->win_netmap_handle;
446 		}
447 	}
448 	return NULL;
449 }
450 
451 /*
452  * we need to wrap ioctl and mmap, at least for the netmap file descriptors
453  */
454 
455 /*
456  * use this function only from netmap_user.h internal functions
457  * same as ioctl, returns 0 on success and -1 on error
458  */
459 static int
460 win_nm_ioctl_internal(HANDLE h, int32_t ctlCode, void *arg)
461 {
462 	DWORD bReturn = 0, szIn, szOut;
463 	BOOL ioctlReturnStatus;
464 	void *inParam = arg, *outParam = arg;
465 
466 	switch (ctlCode) {
467 	case NETMAP_POLL:
468 		szIn = sizeof(POLL_REQUEST_DATA);
469 		szOut = sizeof(POLL_REQUEST_DATA);
470 		break;
471 	case NETMAP_MMAP:
472 		szIn = 0;
473 		szOut = sizeof(void*);
474 		inParam = NULL; /* nothing on input */
475 		break;
476 	case NIOCTXSYNC:
477 	case NIOCRXSYNC:
478 		szIn = 0;
479 		szOut = 0;
480 		break;
481 	case NIOCREGIF:
482 		szIn = sizeof(struct nmreq);
483 		szOut = sizeof(struct nmreq);
484 		break;
485 	case NIOCCONFIG:
486 		D("unsupported NIOCCONFIG!");
487 		return -1;
488 
489 	default: /* a regular ioctl */
490 		D("invalid ioctl %x on netmap fd", ctlCode);
491 		return -1;
492 	}
493 
494 	ioctlReturnStatus = DeviceIoControl(h,
495 				ctlCode, inParam, szIn,
496 				outParam, szOut,
497 				&bReturn, NULL);
498 	// XXX note windows returns 0 on error or async call, 1 on success
499 	// we could call GetLastError() to figure out what happened
500 	return ioctlReturnStatus ? 0 : -1;
501 }
502 
503 /*
504  * this function is what must be called from user-space programs
505  * same as ioctl, returns 0 on success and -1 on error
506  */
507 static int
508 win_nm_ioctl(int fd, int32_t ctlCode, void *arg)
509 {
510 	HANDLE h = win_get_netmap_handle(fd);
511 
512 	if (h == NULL) {
513 		return ioctl(fd, ctlCode, arg);
514 	} else {
515 		return win_nm_ioctl_internal(h, ctlCode, arg);
516 	}
517 }
518 
519 #define ioctl win_nm_ioctl /* from now on, within this file ... */
520 
521 /*
522  * We cannot use the native mmap on windows
523  * The only parameter used is "fd", the other ones are just declared to
524  * make this signature comparable to the FreeBSD/Linux one
525  */
526 static void *
527 win32_mmap_emulated(void *addr, size_t length, int prot, int flags, int fd, int32_t offset)
528 {
529 	HANDLE h = win_get_netmap_handle(fd);
530 
531 	if (h == NULL) {
532 		return mmap(addr, length, prot, flags, fd, offset);
533 	} else {
534 		MEMORY_ENTRY ret;
535 
536 		return win_nm_ioctl_internal(h, NETMAP_MMAP, &ret) ?
537 			NULL : ret.pUsermodeVirtualAddress;
538 	}
539 }
540 
541 #define mmap win32_mmap_emulated
542 
543 #include <sys/poll.h> /* XXX needed to use the structure pollfd */
544 
545 static int
546 win_nm_poll(struct pollfd *fds, int nfds, int timeout)
547 {
548 	HANDLE h;
549 
550 	if (nfds != 1 || fds == NULL || (h = win_get_netmap_handle(fds->fd)) == NULL) {;
551 		return poll(fds, nfds, timeout);
552 	} else {
553 		POLL_REQUEST_DATA prd;
554 
555 		prd.timeout = timeout;
556 		prd.events = fds->events;
557 
558 		win_nm_ioctl_internal(h, NETMAP_POLL, &prd);
559 		if ((prd.revents == POLLERR) || (prd.revents == STATUS_TIMEOUT)) {
560 			return -1;
561 		}
562 		return 1;
563 	}
564 }
565 
566 #define poll win_nm_poll
567 
568 static int
569 win_nm_open(char* pathname, int flags)
570 {
571 
572 	if (strcmp(pathname, NETMAP_DEVICE_NAME) == 0) {
573 		int fd = open(NETMAP_DEVICE_NAME, O_RDWR);
574 		if (fd < 0) {
575 			return -1;
576 		}
577 
578 		win_insert_fd_record(fd);
579 		return fd;
580 	} else {
581 		return open(pathname, flags);
582 	}
583 }
584 
585 #define open win_nm_open
586 
587 static int
588 win_nm_close(int fd)
589 {
590 	if (fd != -1) {
591 		close(fd);
592 		if (win_get_netmap_handle(fd) != NULL) {
593 			win_remove_fd_record(fd);
594 		}
595 	}
596 	return 0;
597 }
598 
599 #define close win_nm_close
600 
601 #endif /* _WIN32 */
602 
603 static int
604 nm_is_identifier(const char *s, const char *e)
605 {
606 	for (; s != e; s++) {
607 		if (!isalnum(*s) && *s != '_') {
608 			return 0;
609 		}
610 	}
611 
612 	return 1;
613 }
614 
615 #define MAXERRMSG 80
616 static int
617 nm_parse(const char *ifname, struct nm_desc *d, char *err)
618 {
619 	int is_vale;
620 	const char *port = NULL;
621 	const char *vpname = NULL;
622 	u_int namelen;
623 	uint32_t nr_ringid = 0, nr_flags;
624 	char errmsg[MAXERRMSG] = "";
625 	long num;
626 	uint16_t nr_arg2 = 0;
627 	enum { P_START, P_RNGSFXOK, P_GETNUM, P_FLAGS, P_FLAGSOK, P_MEMID } p_state;
628 
629 	errno = 0;
630 
631 	is_vale = (ifname[0] == 'v');
632 	if (is_vale) {
633 		port = index(ifname, ':');
634 		if (port == NULL) {
635 			snprintf(errmsg, MAXERRMSG,
636 				 "missing ':' in vale name");
637 			goto fail;
638 		}
639 
640 		if (!nm_is_identifier(ifname + 4, port)) {
641 			snprintf(errmsg, MAXERRMSG, "invalid bridge name");
642 			goto fail;
643 		}
644 
645 		vpname = ++port;
646 	} else {
647 		ifname += 7;
648 		port = ifname;
649 	}
650 
651 	/* scan for a separator */
652 	for (; *port && !index("-*^{}/@", *port); port++)
653 		;
654 
655 	if (is_vale && !nm_is_identifier(vpname, port)) {
656 		snprintf(errmsg, MAXERRMSG, "invalid bridge port name");
657 		goto fail;
658 	}
659 
660 	namelen = port - ifname;
661 	if (namelen >= sizeof(d->req.nr_name)) {
662 		snprintf(errmsg, MAXERRMSG, "name too long");
663 		goto fail;
664 	}
665 	memcpy(d->req.nr_name, ifname, namelen);
666 	d->req.nr_name[namelen] = '\0';
667 
668 	p_state = P_START;
669 	nr_flags = NR_REG_ALL_NIC; /* default for no suffix */
670 	while (*port) {
671 		switch (p_state) {
672 		case P_START:
673 			switch (*port) {
674 			case '^': /* only SW ring */
675 				nr_flags = NR_REG_SW;
676 				p_state = P_RNGSFXOK;
677 				break;
678 			case '*': /* NIC and SW */
679 				nr_flags = NR_REG_NIC_SW;
680 				p_state = P_RNGSFXOK;
681 				break;
682 			case '-': /* one NIC ring pair */
683 				nr_flags = NR_REG_ONE_NIC;
684 				p_state = P_GETNUM;
685 				break;
686 			case '{': /* pipe (master endpoint) */
687 				nr_flags = NR_REG_PIPE_MASTER;
688 				p_state = P_GETNUM;
689 				break;
690 			case '}': /* pipe (slave endoint) */
691 				nr_flags = NR_REG_PIPE_SLAVE;
692 				p_state = P_GETNUM;
693 				break;
694 			case '/': /* start of flags */
695 				p_state = P_FLAGS;
696 				break;
697 			case '@': /* start of memid */
698 				p_state = P_MEMID;
699 				break;
700 			default:
701 				snprintf(errmsg, MAXERRMSG, "unknown modifier: '%c'", *port);
702 				goto fail;
703 			}
704 			port++;
705 			break;
706 		case P_RNGSFXOK:
707 			switch (*port) {
708 			case '/':
709 				p_state = P_FLAGS;
710 				break;
711 			case '@':
712 				p_state = P_MEMID;
713 				break;
714 			default:
715 				snprintf(errmsg, MAXERRMSG, "unexpected character: '%c'", *port);
716 				goto fail;
717 			}
718 			port++;
719 			break;
720 		case P_GETNUM:
721 			num = strtol(port, (char **)&port, 10);
722 			if (num < 0 || num >= NETMAP_RING_MASK) {
723 				snprintf(errmsg, MAXERRMSG, "'%ld' out of range [0, %d)",
724 						num, NETMAP_RING_MASK);
725 				goto fail;
726 			}
727 			nr_ringid = num & NETMAP_RING_MASK;
728 			p_state = P_RNGSFXOK;
729 			break;
730 		case P_FLAGS:
731 		case P_FLAGSOK:
732 			if (*port == '@') {
733 				port++;
734 				p_state = P_MEMID;
735 				break;
736 			}
737 			switch (*port) {
738 			case 'x':
739 				nr_flags |= NR_EXCLUSIVE;
740 				break;
741 			case 'z':
742 				nr_flags |= NR_ZCOPY_MON;
743 				break;
744 			case 't':
745 				nr_flags |= NR_MONITOR_TX;
746 				break;
747 			case 'r':
748 				nr_flags |= NR_MONITOR_RX;
749 				break;
750 			case 'R':
751 				nr_flags |= NR_RX_RINGS_ONLY;
752 				break;
753 			case 'T':
754 				nr_flags |= NR_TX_RINGS_ONLY;
755 				break;
756 			default:
757 				snprintf(errmsg, MAXERRMSG, "unrecognized flag: '%c'", *port);
758 				goto fail;
759 			}
760 			port++;
761 			p_state = P_FLAGSOK;
762 			break;
763 		case P_MEMID:
764 			if (nr_arg2 != 0) {
765 				snprintf(errmsg, MAXERRMSG, "double setting of memid");
766 				goto fail;
767 			}
768 			num = strtol(port, (char **)&port, 10);
769 			if (num <= 0) {
770 				snprintf(errmsg, MAXERRMSG, "invalid memid %ld, must be >0", num);
771 				goto fail;
772 			}
773 			nr_arg2 = num;
774 			p_state = P_RNGSFXOK;
775 			break;
776 		}
777 	}
778 	if (p_state != P_START && p_state != P_RNGSFXOK && p_state != P_FLAGSOK) {
779 		snprintf(errmsg, MAXERRMSG, "unexpected end of port name");
780 		goto fail;
781 	}
782 	ND("flags: %s %s %s %s",
783 			(nr_flags & NR_EXCLUSIVE) ? "EXCLUSIVE" : "",
784 			(nr_flags & NR_ZCOPY_MON) ? "ZCOPY_MON" : "",
785 			(nr_flags & NR_MONITOR_TX) ? "MONITOR_TX" : "",
786 			(nr_flags & NR_MONITOR_RX) ? "MONITOR_RX" : "");
787 
788 	d->req.nr_flags |= nr_flags;
789 	d->req.nr_ringid |= nr_ringid;
790 	d->req.nr_arg2 = nr_arg2;
791 
792 	d->self = d;
793 
794 	return 0;
795 fail:
796 	if (!errno)
797 		errno = EINVAL;
798 	if (err)
799 		strncpy(err, errmsg, MAXERRMSG);
800 	return -1;
801 }
802 
803 /*
804  * Try to open, return descriptor if successful, NULL otherwise.
805  * An invalid netmap name will return errno = 0;
806  * You can pass a pointer to a pre-filled nm_desc to add special
807  * parameters. Flags is used as follows
808  * NM_OPEN_NO_MMAP	use the memory from arg, only XXX avoid mmap
809  *			if the nr_arg2 (memory block) matches.
810  * NM_OPEN_ARG1		use req.nr_arg1 from arg
811  * NM_OPEN_ARG2		use req.nr_arg2 from arg
812  * NM_OPEN_RING_CFG	user ring config from arg
813  */
814 static struct nm_desc *
815 nm_open(const char *ifname, const struct nmreq *req,
816 	uint64_t new_flags, const struct nm_desc *arg)
817 {
818 	struct nm_desc *d = NULL;
819 	const struct nm_desc *parent = arg;
820 	char errmsg[MAXERRMSG] = "";
821 	uint32_t nr_reg;
822 
823 	if (strncmp(ifname, "netmap:", 7) &&
824 			strncmp(ifname, NM_BDG_NAME, strlen(NM_BDG_NAME))) {
825 		errno = 0; /* name not recognised, not an error */
826 		return NULL;
827 	}
828 
829 	d = (struct nm_desc *)calloc(1, sizeof(*d));
830 	if (d == NULL) {
831 		snprintf(errmsg, MAXERRMSG, "nm_desc alloc failure");
832 		errno = ENOMEM;
833 		return NULL;
834 	}
835 	d->self = d;	/* set this early so nm_close() works */
836 	d->fd = open(NETMAP_DEVICE_NAME, O_RDWR);
837 	if (d->fd < 0) {
838 		snprintf(errmsg, MAXERRMSG, "cannot open /dev/netmap: %s", strerror(errno));
839 		goto fail;
840 	}
841 
842 	if (req)
843 		d->req = *req;
844 
845 	if (!(new_flags & NM_OPEN_IFNAME)) {
846 		if (nm_parse(ifname, d, errmsg) < 0)
847 			goto fail;
848 	}
849 
850 	d->req.nr_version = NETMAP_API;
851 	d->req.nr_ringid &= NETMAP_RING_MASK;
852 
853 	/* optionally import info from parent */
854 	if (IS_NETMAP_DESC(parent) && new_flags) {
855 		if (new_flags & NM_OPEN_ARG1)
856 			D("overriding ARG1 %d", parent->req.nr_arg1);
857 		d->req.nr_arg1 = new_flags & NM_OPEN_ARG1 ?
858 			parent->req.nr_arg1 : 4;
859 		if (new_flags & NM_OPEN_ARG2) {
860 			D("overriding ARG2 %d", parent->req.nr_arg2);
861 			d->req.nr_arg2 =  parent->req.nr_arg2;
862 		}
863 		if (new_flags & NM_OPEN_ARG3)
864 			D("overriding ARG3 %d", parent->req.nr_arg3);
865 		d->req.nr_arg3 = new_flags & NM_OPEN_ARG3 ?
866 			parent->req.nr_arg3 : 0;
867 		if (new_flags & NM_OPEN_RING_CFG) {
868 			D("overriding RING_CFG");
869 			d->req.nr_tx_slots = parent->req.nr_tx_slots;
870 			d->req.nr_rx_slots = parent->req.nr_rx_slots;
871 			d->req.nr_tx_rings = parent->req.nr_tx_rings;
872 			d->req.nr_rx_rings = parent->req.nr_rx_rings;
873 		}
874 		if (new_flags & NM_OPEN_IFNAME) {
875 			D("overriding ifname %s ringid 0x%x flags 0x%x",
876 				parent->req.nr_name, parent->req.nr_ringid,
877 				parent->req.nr_flags);
878 			memcpy(d->req.nr_name, parent->req.nr_name,
879 				sizeof(d->req.nr_name));
880 			d->req.nr_ringid = parent->req.nr_ringid;
881 			d->req.nr_flags = parent->req.nr_flags;
882 		}
883 	}
884 	/* add the *XPOLL flags */
885 	d->req.nr_ringid |= new_flags & (NETMAP_NO_TX_POLL | NETMAP_DO_RX_POLL);
886 
887 	if (ioctl(d->fd, NIOCREGIF, &d->req)) {
888 		snprintf(errmsg, MAXERRMSG, "NIOCREGIF failed: %s", strerror(errno));
889 		goto fail;
890 	}
891 
892 	nr_reg = d->req.nr_flags & NR_REG_MASK;
893 
894 	if (nr_reg == NR_REG_SW) { /* host stack */
895 		d->first_tx_ring = d->last_tx_ring = d->req.nr_tx_rings;
896 		d->first_rx_ring = d->last_rx_ring = d->req.nr_rx_rings;
897 	} else if (nr_reg ==  NR_REG_ALL_NIC) { /* only nic */
898 		d->first_tx_ring = 0;
899 		d->first_rx_ring = 0;
900 		d->last_tx_ring = d->req.nr_tx_rings - 1;
901 		d->last_rx_ring = d->req.nr_rx_rings - 1;
902 	} else if (nr_reg ==  NR_REG_NIC_SW) {
903 		d->first_tx_ring = 0;
904 		d->first_rx_ring = 0;
905 		d->last_tx_ring = d->req.nr_tx_rings;
906 		d->last_rx_ring = d->req.nr_rx_rings;
907 	} else if (nr_reg == NR_REG_ONE_NIC) {
908 		/* XXX check validity */
909 		d->first_tx_ring = d->last_tx_ring =
910 		d->first_rx_ring = d->last_rx_ring = d->req.nr_ringid & NETMAP_RING_MASK;
911 	} else { /* pipes */
912 		d->first_tx_ring = d->last_tx_ring = 0;
913 		d->first_rx_ring = d->last_rx_ring = 0;
914 	}
915 
916         /* if parent is defined, do nm_mmap() even if NM_OPEN_NO_MMAP is set */
917 	if ((!(new_flags & NM_OPEN_NO_MMAP) || parent) && nm_mmap(d, parent)) {
918 	        snprintf(errmsg, MAXERRMSG, "mmap failed: %s", strerror(errno));
919 		goto fail;
920 	}
921 
922 
923 #ifdef DEBUG_NETMAP_USER
924     { /* debugging code */
925 	int i;
926 
927 	D("%s tx %d .. %d %d rx %d .. %d %d", ifname,
928 		d->first_tx_ring, d->last_tx_ring, d->req.nr_tx_rings,
929                 d->first_rx_ring, d->last_rx_ring, d->req.nr_rx_rings);
930 	for (i = 0; i <= d->req.nr_tx_rings; i++) {
931 		struct netmap_ring *r = NETMAP_TXRING(d->nifp, i);
932 		D("TX%d %p h %d c %d t %d", i, r, r->head, r->cur, r->tail);
933 	}
934 	for (i = 0; i <= d->req.nr_rx_rings; i++) {
935 		struct netmap_ring *r = NETMAP_RXRING(d->nifp, i);
936 		D("RX%d %p h %d c %d t %d", i, r, r->head, r->cur, r->tail);
937 	}
938     }
939 #endif /* debugging */
940 
941 	d->cur_tx_ring = d->first_tx_ring;
942 	d->cur_rx_ring = d->first_rx_ring;
943 	return d;
944 
945 fail:
946 	nm_close(d);
947 	if (errmsg[0])
948 		D("%s %s", errmsg, ifname);
949 	if (errno == 0)
950 		errno = EINVAL;
951 	return NULL;
952 }
953 
954 
955 static int
956 nm_close(struct nm_desc *d)
957 {
958 	/*
959 	 * ugly trick to avoid unused warnings
960 	 */
961 	static void *__xxzt[] __attribute__ ((unused))  =
962 		{ (void *)nm_open, (void *)nm_inject,
963 		  (void *)nm_dispatch, (void *)nm_nextpkt } ;
964 
965 	if (d == NULL || d->self != d)
966 		return EINVAL;
967 	if (d->done_mmap && d->mem)
968 		munmap(d->mem, d->memsize);
969 	if (d->fd != -1) {
970 		close(d->fd);
971 	}
972 
973 	bzero(d, sizeof(*d));
974 	free(d);
975 	return 0;
976 }
977 
978 
979 static int
980 nm_mmap(struct nm_desc *d, const struct nm_desc *parent)
981 {
982 	//XXX TODO: check if mmap is already done
983 
984 	if (IS_NETMAP_DESC(parent) && parent->mem &&
985 	    parent->req.nr_arg2 == d->req.nr_arg2) {
986 		/* do not mmap, inherit from parent */
987 		D("do not mmap, inherit from parent");
988 		d->memsize = parent->memsize;
989 		d->mem = parent->mem;
990 	} else {
991 		/* XXX TODO: check if memsize is too large (or there is overflow) */
992 		d->memsize = d->req.nr_memsize;
993 		d->mem = mmap(0, d->memsize, PROT_WRITE | PROT_READ, MAP_SHARED,
994 				d->fd, 0);
995 		if (d->mem == MAP_FAILED) {
996 			goto fail;
997 		}
998 		d->done_mmap = 1;
999 	}
1000 	{
1001 		struct netmap_if *nifp = NETMAP_IF(d->mem, d->req.nr_offset);
1002 		struct netmap_ring *r = NETMAP_RXRING(nifp, d->first_rx_ring);
1003 		if ((void *)r == (void *)nifp) {
1004 			/* the descriptor is open for TX only */
1005 			r = NETMAP_TXRING(nifp, d->first_tx_ring);
1006 		}
1007 
1008 		*(struct netmap_if **)(uintptr_t)&(d->nifp) = nifp;
1009 		*(struct netmap_ring **)(uintptr_t)&d->some_ring = r;
1010 		*(void **)(uintptr_t)&d->buf_start = NETMAP_BUF(r, 0);
1011 		*(void **)(uintptr_t)&d->buf_end =
1012 			(char *)d->mem + d->memsize;
1013 	}
1014 
1015 	return 0;
1016 
1017 fail:
1018 	return EINVAL;
1019 }
1020 
1021 /*
1022  * Same prototype as pcap_inject(), only need to cast.
1023  */
1024 static int
1025 nm_inject(struct nm_desc *d, const void *buf, size_t size)
1026 {
1027 	u_int c, n = d->last_tx_ring - d->first_tx_ring + 1,
1028 		ri = d->cur_tx_ring;
1029 
1030 	for (c = 0; c < n ; c++, ri++) {
1031 		/* compute current ring to use */
1032 		struct netmap_ring *ring;
1033 		uint32_t i, j, idx;
1034 		size_t rem;
1035 
1036 		if (ri > d->last_tx_ring)
1037 			ri = d->first_tx_ring;
1038 		ring = NETMAP_TXRING(d->nifp, ri);
1039 		rem = size;
1040 		j = ring->cur;
1041 		while (rem > ring->nr_buf_size && j != ring->tail) {
1042 			rem -= ring->nr_buf_size;
1043 			j = nm_ring_next(ring, j);
1044 		}
1045 		if (j == ring->tail && rem > 0)
1046 			continue;
1047 		i = ring->cur;
1048 		while (i != j) {
1049 			idx = ring->slot[i].buf_idx;
1050 			ring->slot[i].len = ring->nr_buf_size;
1051 			ring->slot[i].flags = NS_MOREFRAG;
1052 			nm_pkt_copy(buf, NETMAP_BUF(ring, idx), ring->nr_buf_size);
1053 			i = nm_ring_next(ring, i);
1054 			buf = (char *)buf + ring->nr_buf_size;
1055 		}
1056 		idx = ring->slot[i].buf_idx;
1057 		ring->slot[i].len = rem;
1058 		ring->slot[i].flags = 0;
1059 		nm_pkt_copy(buf, NETMAP_BUF(ring, idx), rem);
1060 		ring->head = ring->cur = nm_ring_next(ring, i);
1061 		d->cur_tx_ring = ri;
1062 		return size;
1063 	}
1064 	return 0; /* fail */
1065 }
1066 
1067 
1068 /*
1069  * Same prototype as pcap_dispatch(), only need to cast.
1070  */
1071 static int
1072 nm_dispatch(struct nm_desc *d, int cnt, nm_cb_t cb, u_char *arg)
1073 {
1074 	int n = d->last_rx_ring - d->first_rx_ring + 1;
1075 	int c, got = 0, ri = d->cur_rx_ring;
1076 	d->hdr.buf = NULL;
1077 	d->hdr.flags = NM_MORE_PKTS;
1078 	d->hdr.d = d;
1079 
1080 	if (cnt == 0)
1081 		cnt = -1;
1082 	/* cnt == -1 means infinite, but rings have a finite amount
1083 	 * of buffers and the int is large enough that we never wrap,
1084 	 * so we can omit checking for -1
1085 	 */
1086 	for (c=0; c < n && cnt != got; c++, ri++) {
1087 		/* compute current ring to use */
1088 		struct netmap_ring *ring;
1089 
1090 		if (ri > d->last_rx_ring)
1091 			ri = d->first_rx_ring;
1092 		ring = NETMAP_RXRING(d->nifp, ri);
1093 		for ( ; !nm_ring_empty(ring) && cnt != got; got++) {
1094 			u_int idx, i;
1095 			u_char *oldbuf;
1096 			struct netmap_slot *slot;
1097 			if (d->hdr.buf) { /* from previous round */
1098 				cb(arg, &d->hdr, d->hdr.buf);
1099 			}
1100 			i = ring->cur;
1101 			slot = &ring->slot[i];
1102 			idx = slot->buf_idx;
1103 			/* d->cur_rx_ring doesn't change inside this loop, but
1104 			 * set it here, so it reflects d->hdr.buf's ring */
1105 			d->cur_rx_ring = ri;
1106 			d->hdr.slot = slot;
1107 			oldbuf = d->hdr.buf = (u_char *)NETMAP_BUF(ring, idx);
1108 			// __builtin_prefetch(buf);
1109 			d->hdr.len = d->hdr.caplen = slot->len;
1110 			while (slot->flags & NS_MOREFRAG) {
1111 				u_char *nbuf;
1112 				u_int oldlen = slot->len;
1113 				i = nm_ring_next(ring, i);
1114 				slot = &ring->slot[i];
1115 				d->hdr.len += slot->len;
1116 				nbuf = (u_char *)NETMAP_BUF(ring, slot->buf_idx);
1117 				if (oldbuf != NULL && nbuf - oldbuf == ring->nr_buf_size &&
1118 						oldlen == ring->nr_buf_size) {
1119 					d->hdr.caplen += slot->len;
1120 					oldbuf = nbuf;
1121 				} else {
1122 					oldbuf = NULL;
1123 				}
1124 			}
1125 			d->hdr.ts = ring->ts;
1126 			ring->head = ring->cur = nm_ring_next(ring, i);
1127 		}
1128 	}
1129 	if (d->hdr.buf) { /* from previous round */
1130 		d->hdr.flags = 0;
1131 		cb(arg, &d->hdr, d->hdr.buf);
1132 	}
1133 	return got;
1134 }
1135 
1136 static u_char *
1137 nm_nextpkt(struct nm_desc *d, struct nm_pkthdr *hdr)
1138 {
1139 	int ri = d->cur_rx_ring;
1140 
1141 	do {
1142 		/* compute current ring to use */
1143 		struct netmap_ring *ring = NETMAP_RXRING(d->nifp, ri);
1144 		if (!nm_ring_empty(ring)) {
1145 			u_int i = ring->cur;
1146 			u_int idx = ring->slot[i].buf_idx;
1147 			u_char *buf = (u_char *)NETMAP_BUF(ring, idx);
1148 
1149 			// __builtin_prefetch(buf);
1150 			hdr->ts = ring->ts;
1151 			hdr->len = hdr->caplen = ring->slot[i].len;
1152 			ring->cur = nm_ring_next(ring, i);
1153 			/* we could postpone advancing head if we want
1154 			 * to hold the buffer. This can be supported in
1155 			 * the future.
1156 			 */
1157 			ring->head = ring->cur;
1158 			d->cur_rx_ring = ri;
1159 			return buf;
1160 		}
1161 		ri++;
1162 		if (ri > d->last_rx_ring)
1163 			ri = d->first_rx_ring;
1164 	} while (ri != d->cur_rx_ring);
1165 	return NULL; /* nothing found */
1166 }
1167 
1168 #endif /* !HAVE_NETMAP_WITH_LIBS */
1169 
1170 #endif /* NETMAP_WITH_LIBS */
1171 
1172 #endif /* _NET_NETMAP_USER_H_ */
1173