xref: /freebsd-14.2/sys/net/netmap_user.h (revision f0ea3689)
168b8534bSLuigi Rizzo /*
217885a7bSLuigi Rizzo  * Copyright (C) 2011-2014 Universita` di Pisa. All rights reserved.
368b8534bSLuigi Rizzo  *
468b8534bSLuigi Rizzo  * Redistribution and use in source and binary forms, with or without
5f9790aebSLuigi Rizzo  * modification, are permitted provided that the following conditions
6f9790aebSLuigi Rizzo  * are met:
768b8534bSLuigi Rizzo  *
868b8534bSLuigi Rizzo  *   1. Redistributions of source code must retain the above copyright
968b8534bSLuigi Rizzo  *      notice, this list of conditions and the following disclaimer.
1068b8534bSLuigi Rizzo  *   2. Redistributions in binary form must reproduce the above copyright
1168b8534bSLuigi Rizzo  *      notice, this list of conditions and the following disclaimer in the
12f9790aebSLuigi Rizzo  *      documentation and/or other materials provided with the distribution.
1368b8534bSLuigi Rizzo  *
14f9790aebSLuigi Rizzo  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
1568b8534bSLuigi Rizzo  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16f9790aebSLuigi Rizzo  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17f9790aebSLuigi Rizzo  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18f9790aebSLuigi Rizzo  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19f9790aebSLuigi Rizzo  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20f9790aebSLuigi Rizzo  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21f9790aebSLuigi Rizzo  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22f9790aebSLuigi Rizzo  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23f9790aebSLuigi Rizzo  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24f9790aebSLuigi Rizzo  * SUCH DAMAGE.
2568b8534bSLuigi Rizzo  */
2668b8534bSLuigi Rizzo 
2768b8534bSLuigi Rizzo /*
2868b8534bSLuigi Rizzo  * $FreeBSD$
2968b8534bSLuigi Rizzo  *
3017885a7bSLuigi Rizzo  * Functions and macros to manipulate netmap structures and packets
3117885a7bSLuigi Rizzo  * in userspace. See netmap(4) for more information.
3268b8534bSLuigi Rizzo  *
3364ae02c3SLuigi Rizzo  * The address of the struct netmap_if, say nifp, is computed from the
3464ae02c3SLuigi Rizzo  * value returned from ioctl(.., NIOCREG, ...) and the mmap region:
3568b8534bSLuigi Rizzo  *	ioctl(fd, NIOCREG, &req);
3668b8534bSLuigi Rizzo  *	mem = mmap(0, ... );
3768b8534bSLuigi Rizzo  *	nifp = NETMAP_IF(mem, req.nr_nifp);
3868b8534bSLuigi Rizzo  *		(so simple, we could just do it manually)
3968b8534bSLuigi Rizzo  *
4068b8534bSLuigi Rizzo  * From there:
4168b8534bSLuigi Rizzo  *	struct netmap_ring *NETMAP_TXRING(nifp, index)
4268b8534bSLuigi Rizzo  *	struct netmap_ring *NETMAP_RXRING(nifp, index)
4368b8534bSLuigi Rizzo  *		we can access ring->nr_cur, ring->nr_avail, ring->nr_flags
4468b8534bSLuigi Rizzo  *
4568b8534bSLuigi Rizzo  *	ring->slot[i] gives us the i-th slot (we can access
4617885a7bSLuigi Rizzo  *		directly len, flags, buf_idx)
4768b8534bSLuigi Rizzo  *
4827892e02SLuigi Rizzo  *	char *buf = NETMAP_BUF(ring, x) returns a pointer to
4927892e02SLuigi Rizzo  *		the buffer numbered x
5068b8534bSLuigi Rizzo  *
5117885a7bSLuigi Rizzo  * All ring indexes (head, cur, tail) should always move forward.
5217885a7bSLuigi Rizzo  * To compute the next index in a circular ring you can use
5317885a7bSLuigi Rizzo  *	i = nm_ring_next(ring, i);
54f9790aebSLuigi Rizzo  *
55f9790aebSLuigi Rizzo  * To ease porting apps from pcap to netmap we supply a few fuctions
5617885a7bSLuigi Rizzo  * that can be called to open, close, read and write on netmap in a way
5717885a7bSLuigi Rizzo  * similar to libpcap. Note that the read/write function depend on
5817885a7bSLuigi Rizzo  * an ioctl()/select()/poll() being issued to refill rings or push
5917885a7bSLuigi Rizzo  * packets out.
60f9790aebSLuigi Rizzo  *
61f9790aebSLuigi Rizzo  * In order to use these, include #define NETMAP_WITH_LIBS
62f9790aebSLuigi Rizzo  * in the source file that invokes these functions.
6368b8534bSLuigi Rizzo  */
6468b8534bSLuigi Rizzo 
6568b8534bSLuigi Rizzo #ifndef _NET_NETMAP_USER_H_
6668b8534bSLuigi Rizzo #define _NET_NETMAP_USER_H_
6768b8534bSLuigi Rizzo 
68f9790aebSLuigi Rizzo #include <stdint.h>
69*f0ea3689SLuigi Rizzo #include <sys/socket.h>		/* apple needs sockaddr */
70f9790aebSLuigi Rizzo #include <net/if.h>		/* IFNAMSIZ */
7117885a7bSLuigi Rizzo 
7217885a7bSLuigi Rizzo #ifndef likely
7317885a7bSLuigi Rizzo #define likely(x)	__builtin_expect(!!(x), 1)
7417885a7bSLuigi Rizzo #define unlikely(x)	__builtin_expect(!!(x), 0)
7517885a7bSLuigi Rizzo #endif /* likely and unlikely */
7617885a7bSLuigi Rizzo 
77f9790aebSLuigi Rizzo #include <net/netmap.h>
78f9790aebSLuigi Rizzo 
7917885a7bSLuigi Rizzo /* helper macro */
80104d9fc7SEd Maste #define _NETMAP_OFFSET(type, ptr, offset) \
81104d9fc7SEd Maste 	((type)(void *)((char *)(ptr) + (offset)))
8268b8534bSLuigi Rizzo 
8317885a7bSLuigi Rizzo #define NETMAP_IF(_base, _ofs)	_NETMAP_OFFSET(struct netmap_if *, _base, _ofs)
8468b8534bSLuigi Rizzo 
85104d9fc7SEd Maste #define NETMAP_TXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *, \
86104d9fc7SEd Maste 	nifp, (nifp)->ring_ofs[index] )
87104d9fc7SEd Maste 
88104d9fc7SEd Maste #define NETMAP_RXRING(nifp, index) _NETMAP_OFFSET(struct netmap_ring *,	\
89104d9fc7SEd Maste 	nifp, (nifp)->ring_ofs[index + (nifp)->ni_tx_rings + 1] )
9068b8534bSLuigi Rizzo 
9168b8534bSLuigi Rizzo #define NETMAP_BUF(ring, index)				\
925819da83SLuigi Rizzo 	((char *)(ring) + (ring)->buf_ofs + ((index)*(ring)->nr_buf_size))
9368b8534bSLuigi Rizzo 
9464ae02c3SLuigi Rizzo #define NETMAP_BUF_IDX(ring, buf)			\
9564ae02c3SLuigi Rizzo 	( ((char *)(buf) - ((char *)(ring) + (ring)->buf_ofs) ) / \
96a7250582SLuigi Rizzo 		(ring)->nr_buf_size )
9764ae02c3SLuigi Rizzo 
9868b8534bSLuigi Rizzo 
9917885a7bSLuigi Rizzo static inline uint32_t
10017885a7bSLuigi Rizzo nm_ring_next(struct netmap_ring *r, uint32_t i)
10117885a7bSLuigi Rizzo {
10217885a7bSLuigi Rizzo 	return ( unlikely(i + 1 == r->num_slots) ? 0 : i + 1);
10317885a7bSLuigi Rizzo }
10417885a7bSLuigi Rizzo 
105a7250582SLuigi Rizzo 
10668b8534bSLuigi Rizzo /*
10717885a7bSLuigi Rizzo  * Return 1 if we have pending transmissions in the tx ring.
108*f0ea3689SLuigi Rizzo  * When everything is complete ring->head = ring->tail + 1 (modulo ring size)
10968b8534bSLuigi Rizzo  */
11017885a7bSLuigi Rizzo static inline int
11117885a7bSLuigi Rizzo nm_tx_pending(struct netmap_ring *r)
11217885a7bSLuigi Rizzo {
113*f0ea3689SLuigi Rizzo 	return nm_ring_next(r, r->tail) != r->head;
11417885a7bSLuigi Rizzo }
11517885a7bSLuigi Rizzo 
11617885a7bSLuigi Rizzo 
11717885a7bSLuigi Rizzo static inline uint32_t
11817885a7bSLuigi Rizzo nm_ring_space(struct netmap_ring *ring)
11917885a7bSLuigi Rizzo {
12017885a7bSLuigi Rizzo         int ret = ring->tail - ring->cur;
12117885a7bSLuigi Rizzo         if (ret < 0)
12217885a7bSLuigi Rizzo                 ret += ring->num_slots;
12317885a7bSLuigi Rizzo         return ret;
12417885a7bSLuigi Rizzo }
12517885a7bSLuigi Rizzo 
12668b8534bSLuigi Rizzo 
127f9790aebSLuigi Rizzo #ifdef NETMAP_WITH_LIBS
128f9790aebSLuigi Rizzo /*
129f9790aebSLuigi Rizzo  * Support for simple I/O libraries.
130f9790aebSLuigi Rizzo  * Include other system headers required for compiling this.
131f9790aebSLuigi Rizzo  */
132f9790aebSLuigi Rizzo 
133f9790aebSLuigi Rizzo #ifndef HAVE_NETMAP_WITH_LIBS
134f9790aebSLuigi Rizzo #define HAVE_NETMAP_WITH_LIBS
135f9790aebSLuigi Rizzo 
136f9790aebSLuigi Rizzo #include <sys/time.h>
137f9790aebSLuigi Rizzo #include <sys/mman.h>
138f9790aebSLuigi Rizzo #include <string.h>	/* memset */
139f9790aebSLuigi Rizzo #include <sys/ioctl.h>
140f9790aebSLuigi Rizzo #include <sys/errno.h>	/* EINVAL */
141f9790aebSLuigi Rizzo #include <fcntl.h>	/* O_RDWR */
14217885a7bSLuigi Rizzo #include <unistd.h>	/* close() */
14366015019SLuigi Rizzo #include <signal.h>
14417885a7bSLuigi Rizzo #include <stdlib.h>
145f9790aebSLuigi Rizzo 
146*f0ea3689SLuigi Rizzo #ifndef ND /* debug macros */
147*f0ea3689SLuigi Rizzo /* debug support */
148*f0ea3689SLuigi Rizzo #define ND(_fmt, ...) do {} while(0)
149*f0ea3689SLuigi Rizzo #define D(_fmt, ...)						\
150*f0ea3689SLuigi Rizzo 	do {							\
151*f0ea3689SLuigi Rizzo 		struct timeval t0;				\
152*f0ea3689SLuigi Rizzo 		gettimeofday(&t0, NULL);			\
153*f0ea3689SLuigi Rizzo 		fprintf(stderr, "%03d.%06d %s [%d] " _fmt "\n",	\
154*f0ea3689SLuigi Rizzo 		    (int)(t0.tv_sec % 1000), (int)t0.tv_usec,	\
155*f0ea3689SLuigi Rizzo 		    __FUNCTION__, __LINE__, ##__VA_ARGS__);	\
156*f0ea3689SLuigi Rizzo         } while (0)
157*f0ea3689SLuigi Rizzo 
158*f0ea3689SLuigi Rizzo /* Rate limited version of "D", lps indicates how many per second */
159*f0ea3689SLuigi Rizzo #define RD(lps, format, ...)                                    \
160*f0ea3689SLuigi Rizzo     do {                                                        \
161*f0ea3689SLuigi Rizzo         static int t0, __cnt;                                   \
162*f0ea3689SLuigi Rizzo         struct timeval __xxts;                                  \
163*f0ea3689SLuigi Rizzo         gettimeofday(&__xxts, NULL);                            \
164*f0ea3689SLuigi Rizzo         if (t0 != __xxts.tv_sec) {                              \
165*f0ea3689SLuigi Rizzo             t0 = __xxts.tv_sec;                                 \
166*f0ea3689SLuigi Rizzo             __cnt = 0;                                          \
167*f0ea3689SLuigi Rizzo         }                                                       \
168*f0ea3689SLuigi Rizzo         if (__cnt++ < lps) {                                    \
169*f0ea3689SLuigi Rizzo             D(format, ##__VA_ARGS__);                           \
170*f0ea3689SLuigi Rizzo         }                                                       \
171*f0ea3689SLuigi Rizzo     } while (0)
172*f0ea3689SLuigi Rizzo #endif
173*f0ea3689SLuigi Rizzo 
174*f0ea3689SLuigi Rizzo struct nm_pkthdr {	/* same as pcap_pkthdr */
175f9790aebSLuigi Rizzo 	struct timeval	ts;
176f9790aebSLuigi Rizzo 	uint32_t	caplen;
177f9790aebSLuigi Rizzo 	uint32_t	len;
178f9790aebSLuigi Rizzo };
179f9790aebSLuigi Rizzo 
180*f0ea3689SLuigi Rizzo struct nm_stat {	/* same as pcap_stat	*/
18166015019SLuigi Rizzo 	u_int	ps_recv;
18266015019SLuigi Rizzo 	u_int	ps_drop;
18366015019SLuigi Rizzo 	u_int	ps_ifdrop;
18466015019SLuigi Rizzo #ifdef WIN32
18566015019SLuigi Rizzo 	u_int	bs_capt;
18666015019SLuigi Rizzo #endif /* WIN32 */
18766015019SLuigi Rizzo };
18866015019SLuigi Rizzo 
18966015019SLuigi Rizzo #define NM_ERRBUF_SIZE	512
19066015019SLuigi Rizzo 
191*f0ea3689SLuigi Rizzo struct nm_desc {
192*f0ea3689SLuigi Rizzo 	struct nm_desc *self; /* point to self if netmap. */
193f9790aebSLuigi Rizzo 	int fd;
194f9790aebSLuigi Rizzo 	void *mem;
195f9790aebSLuigi Rizzo 	int memsize;
196*f0ea3689SLuigi Rizzo 	int done_mmap;	/* set if mem is the result of mmap */
197*f0ea3689SLuigi Rizzo 	struct netmap_if * const nifp;
19866015019SLuigi Rizzo 	uint16_t first_tx_ring, last_tx_ring, cur_tx_ring;
19966015019SLuigi Rizzo 	uint16_t first_rx_ring, last_rx_ring, cur_rx_ring;
20066015019SLuigi Rizzo 	struct nmreq req;	/* also contains the nr_name = ifname */
201*f0ea3689SLuigi Rizzo 	struct nm_pkthdr hdr;
20266015019SLuigi Rizzo 
203*f0ea3689SLuigi Rizzo 	/*
204*f0ea3689SLuigi Rizzo 	 * The memory contains netmap_if, rings and then buffers.
205*f0ea3689SLuigi Rizzo 	 * Given a pointer (e.g. to nm_inject) we can compare with
206*f0ea3689SLuigi Rizzo 	 * mem/buf_start/buf_end to tell if it is a buffer or
207*f0ea3689SLuigi Rizzo 	 * some other descriptor in our region.
208*f0ea3689SLuigi Rizzo 	 * We also store a pointer to some ring as it helps in the
209*f0ea3689SLuigi Rizzo 	 * translation from buffer indexes to addresses.
210*f0ea3689SLuigi Rizzo 	 */
211*f0ea3689SLuigi Rizzo 	struct netmap_ring * const some_ring;
212*f0ea3689SLuigi Rizzo 	void * const buf_start;
213*f0ea3689SLuigi Rizzo 	void * const buf_end;
21466015019SLuigi Rizzo 	/* parameters from pcap_open_live */
21566015019SLuigi Rizzo 	int snaplen;
21666015019SLuigi Rizzo 	int promisc;
21766015019SLuigi Rizzo 	int to_ms;
21866015019SLuigi Rizzo 	char *errbuf;
21966015019SLuigi Rizzo 
22066015019SLuigi Rizzo 	/* save flags so we can restore them on close */
22166015019SLuigi Rizzo 	uint32_t if_flags;
22266015019SLuigi Rizzo         uint32_t if_reqcap;
22366015019SLuigi Rizzo         uint32_t if_curcap;
22466015019SLuigi Rizzo 
225*f0ea3689SLuigi Rizzo 	struct nm_stat st;
22666015019SLuigi Rizzo 	char msg[NM_ERRBUF_SIZE];
227f9790aebSLuigi Rizzo };
228f9790aebSLuigi Rizzo 
229f9790aebSLuigi Rizzo /*
230f9790aebSLuigi Rizzo  * when the descriptor is open correctly, d->self == d
231b82b2211SLuigi Rizzo  * Eventually we should also use some magic number.
232f9790aebSLuigi Rizzo  */
233*f0ea3689SLuigi Rizzo #define P2NMD(p)		((struct nm_desc *)(p))
234*f0ea3689SLuigi Rizzo #define IS_NETMAP_DESC(d)	((d) && P2NMD(d)->self == P2NMD(d))
235f9790aebSLuigi Rizzo #define NETMAP_FD(d)		(P2NMD(d)->fd)
236f9790aebSLuigi Rizzo 
23717885a7bSLuigi Rizzo 
23817885a7bSLuigi Rizzo /*
23917885a7bSLuigi Rizzo  * this is a slightly optimized copy routine which rounds
24017885a7bSLuigi Rizzo  * to multiple of 64 bytes and is often faster than dealing
24117885a7bSLuigi Rizzo  * with other odd sizes. We assume there is enough room
24217885a7bSLuigi Rizzo  * in the source and destination buffers.
24317885a7bSLuigi Rizzo  *
24417885a7bSLuigi Rizzo  * XXX only for multiples of 64 bytes, non overlapped.
24517885a7bSLuigi Rizzo  */
24617885a7bSLuigi Rizzo static inline void
247*f0ea3689SLuigi Rizzo nm_pkt_copy(const void *_src, void *_dst, int l)
24817885a7bSLuigi Rizzo {
249b82b2211SLuigi Rizzo 	const uint64_t *src = (const uint64_t *)_src;
250b82b2211SLuigi Rizzo 	uint64_t *dst = (uint64_t *)_dst;
251b82b2211SLuigi Rizzo 
25217885a7bSLuigi Rizzo 	if (unlikely(l >= 1024)) {
25317885a7bSLuigi Rizzo 		memcpy(dst, src, l);
25417885a7bSLuigi Rizzo 		return;
25517885a7bSLuigi Rizzo 	}
25617885a7bSLuigi Rizzo 	for (; likely(l > 0); l-=64) {
25717885a7bSLuigi Rizzo 		*dst++ = *src++;
25817885a7bSLuigi Rizzo 		*dst++ = *src++;
25917885a7bSLuigi Rizzo 		*dst++ = *src++;
26017885a7bSLuigi Rizzo 		*dst++ = *src++;
26117885a7bSLuigi Rizzo 		*dst++ = *src++;
26217885a7bSLuigi Rizzo 		*dst++ = *src++;
26317885a7bSLuigi Rizzo 		*dst++ = *src++;
26417885a7bSLuigi Rizzo 		*dst++ = *src++;
26517885a7bSLuigi Rizzo 	}
26617885a7bSLuigi Rizzo }
26717885a7bSLuigi Rizzo 
26817885a7bSLuigi Rizzo 
269f9790aebSLuigi Rizzo /*
270f9790aebSLuigi Rizzo  * The callback, invoked on each received packet. Same as libpcap
271f9790aebSLuigi Rizzo  */
272*f0ea3689SLuigi Rizzo typedef void (*nm_cb_t)(u_char *, const struct nm_pkthdr *, const u_char *d);
273f9790aebSLuigi Rizzo 
274f9790aebSLuigi Rizzo /*
27517885a7bSLuigi Rizzo  *--- the pcap-like API ---
27617885a7bSLuigi Rizzo  *
27717885a7bSLuigi Rizzo  * nm_open() opens a file descriptor, binds to a port and maps memory.
27817885a7bSLuigi Rizzo  *
27917885a7bSLuigi Rizzo  * ifname	(netmap:foo or vale:foo) is the port name
280*f0ea3689SLuigi Rizzo  *		a suffix can indicate the follwing:
281*f0ea3689SLuigi Rizzo  *		^		bind the host (sw) ring pair
282*f0ea3689SLuigi Rizzo  *		*		bind host and NIC ring pairs (transparent)
283*f0ea3689SLuigi Rizzo  *		-NN		bind individual NIC ring pair
284*f0ea3689SLuigi Rizzo  *		{NN		bind master side of pipe NN
285*f0ea3689SLuigi Rizzo  *		}NN		bind slave side of pipe NN
286*f0ea3689SLuigi Rizzo  *
287*f0ea3689SLuigi Rizzo  * req		provides the initial values of nmreq before parsing ifname.
288*f0ea3689SLuigi Rizzo  *		Remember that the ifname parsing will override the ring
289*f0ea3689SLuigi Rizzo  *		number in nm_ringid, and part of nm_flags;
290*f0ea3689SLuigi Rizzo  * flags	special functions, normally 0
291*f0ea3689SLuigi Rizzo  *		indicates which fields of *arg are significant
292*f0ea3689SLuigi Rizzo  * arg		special functions, normally NULL
293*f0ea3689SLuigi Rizzo  *		if passed a netmap_desc with mem != NULL,
294*f0ea3689SLuigi Rizzo  *		use that memory instead of mmap.
295f9790aebSLuigi Rizzo  */
29617885a7bSLuigi Rizzo 
297*f0ea3689SLuigi Rizzo static struct nm_desc *nm_open(const char *ifname, const struct nmreq *req,
298*f0ea3689SLuigi Rizzo 	uint64_t flags, const struct nm_desc *arg);
299*f0ea3689SLuigi Rizzo 
300*f0ea3689SLuigi Rizzo /*
301*f0ea3689SLuigi Rizzo  * nm_open can import some fields from the parent descriptor.
302*f0ea3689SLuigi Rizzo  * These flags control which ones.
303*f0ea3689SLuigi Rizzo  * Also in flags you can specify NETMAP_NO_TX_POLL and NETMAP_DO_RX_POLL,
304*f0ea3689SLuigi Rizzo  * which set the initial value for these flags.
305*f0ea3689SLuigi Rizzo  * Note that the 16 low bits of the flags are reserved for data
306*f0ea3689SLuigi Rizzo  * that may go into the nmreq.
307*f0ea3689SLuigi Rizzo  */
308*f0ea3689SLuigi Rizzo enum {
309*f0ea3689SLuigi Rizzo 	NM_OPEN_NO_MMAP =	0x040000, /* reuse mmap from parent */
310*f0ea3689SLuigi Rizzo 	NM_OPEN_IFNAME =	0x080000, /* nr_name, nr_ringid, nr_flags */
311*f0ea3689SLuigi Rizzo 	NM_OPEN_ARG1 =		0x100000,
312*f0ea3689SLuigi Rizzo 	NM_OPEN_ARG2 =		0x200000,
313*f0ea3689SLuigi Rizzo 	NM_OPEN_ARG3 =		0x400000,
314*f0ea3689SLuigi Rizzo 	NM_OPEN_RING_CFG =	0x800000, /* tx|rx rings|slots */
315*f0ea3689SLuigi Rizzo };
316*f0ea3689SLuigi Rizzo 
317f9790aebSLuigi Rizzo 
318f9790aebSLuigi Rizzo /*
31917885a7bSLuigi Rizzo  * nm_close()	closes and restores the port to its previous state
320f9790aebSLuigi Rizzo  */
32117885a7bSLuigi Rizzo 
322*f0ea3689SLuigi Rizzo static int nm_close(struct nm_desc *);
323f9790aebSLuigi Rizzo 
324f9790aebSLuigi Rizzo /*
32517885a7bSLuigi Rizzo  * nm_inject() is the same as pcap_inject()
32617885a7bSLuigi Rizzo  * nm_dispatch() is the same as pcap_dispatch()
32717885a7bSLuigi Rizzo  * nm_nextpkt() is the same as pcap_next()
328f9790aebSLuigi Rizzo  */
32917885a7bSLuigi Rizzo 
330*f0ea3689SLuigi Rizzo static int nm_inject(struct nm_desc *, const void *, size_t);
331*f0ea3689SLuigi Rizzo static int nm_dispatch(struct nm_desc *, int, nm_cb_t, u_char *);
332*f0ea3689SLuigi Rizzo static u_char *nm_nextpkt(struct nm_desc *, struct nm_pkthdr *);
333f9790aebSLuigi Rizzo 
334f9790aebSLuigi Rizzo 
335f9790aebSLuigi Rizzo /*
336f9790aebSLuigi Rizzo  * Try to open, return descriptor if successful, NULL otherwise.
337f9790aebSLuigi Rizzo  * An invalid netmap name will return errno = 0;
338*f0ea3689SLuigi Rizzo  * You can pass a pointer to a pre-filled nm_desc to add special
339*f0ea3689SLuigi Rizzo  * parameters. Flags is used as follows
340*f0ea3689SLuigi Rizzo  * NM_OPEN_NO_MMAP	use the memory from arg, only
341*f0ea3689SLuigi Rizzo  *			if the nr_arg2 (memory block) matches.
342*f0ea3689SLuigi Rizzo  * NM_OPEN_ARG1		use req.nr_arg1 from arg
343*f0ea3689SLuigi Rizzo  * NM_OPEN_ARG2		use req.nr_arg2 from arg
344*f0ea3689SLuigi Rizzo  * NM_OPEN_RING_CFG	user ring config from arg
345f9790aebSLuigi Rizzo  */
346*f0ea3689SLuigi Rizzo static struct nm_desc *
347*f0ea3689SLuigi Rizzo nm_open(const char *ifname, const struct nmreq *req,
348*f0ea3689SLuigi Rizzo 	uint64_t new_flags, const struct nm_desc *arg)
349f9790aebSLuigi Rizzo {
350*f0ea3689SLuigi Rizzo 	struct nm_desc *d = NULL;
351*f0ea3689SLuigi Rizzo 	const struct nm_desc *parent = arg;
352*f0ea3689SLuigi Rizzo 	u_int namelen;
353*f0ea3689SLuigi Rizzo 	uint32_t nr_ringid = 0, nr_flags;
354*f0ea3689SLuigi Rizzo 	const char *port = NULL;
355*f0ea3689SLuigi Rizzo 	const char *errmsg = NULL;
356f9790aebSLuigi Rizzo 
357f9790aebSLuigi Rizzo 	if (strncmp(ifname, "netmap:", 7) && strncmp(ifname, "vale", 4)) {
358*f0ea3689SLuigi Rizzo 		errno = 0; /* name not recognised, not an error */
359f9790aebSLuigi Rizzo 		return NULL;
360f9790aebSLuigi Rizzo 	}
361f9790aebSLuigi Rizzo 	if (ifname[0] == 'n')
362f9790aebSLuigi Rizzo 		ifname += 7;
363*f0ea3689SLuigi Rizzo 	/* scan for a separator */
364*f0ea3689SLuigi Rizzo 	for (port = ifname; *port && !index("-*^{}", *port); port++)
365*f0ea3689SLuigi Rizzo 		;
36666015019SLuigi Rizzo 	namelen = port - ifname;
367*f0ea3689SLuigi Rizzo 	if (namelen >= sizeof(d->req.nr_name)) {
368*f0ea3689SLuigi Rizzo 		errmsg = "name too long";
369*f0ea3689SLuigi Rizzo 		goto fail;
37066015019SLuigi Rizzo 	}
371*f0ea3689SLuigi Rizzo 	switch (*port) {
372*f0ea3689SLuigi Rizzo 	default:  /* '\0', no suffix */
373*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_ALL_NIC;
374*f0ea3689SLuigi Rizzo 		break;
375*f0ea3689SLuigi Rizzo 	case '-': /* one NIC */
376*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_ONE_NIC;
377*f0ea3689SLuigi Rizzo 		nr_ringid = atoi(port + 1);
378*f0ea3689SLuigi Rizzo 		break;
379*f0ea3689SLuigi Rizzo 	case '*': /* NIC and SW, ignore port */
380*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_NIC_SW;
381*f0ea3689SLuigi Rizzo 		if (port[1]) {
382*f0ea3689SLuigi Rizzo 			errmsg = "invalid port for nic+sw";
383*f0ea3689SLuigi Rizzo 			goto fail;
384*f0ea3689SLuigi Rizzo 		}
385*f0ea3689SLuigi Rizzo 		break;
386*f0ea3689SLuigi Rizzo 	case '^': /* only sw ring */
387*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_SW;
388*f0ea3689SLuigi Rizzo 		if (port[1]) {
389*f0ea3689SLuigi Rizzo 			errmsg = "invalid port for sw ring";
390*f0ea3689SLuigi Rizzo 			goto fail;
391*f0ea3689SLuigi Rizzo 		}
392*f0ea3689SLuigi Rizzo 		break;
393*f0ea3689SLuigi Rizzo 	case '{':
394*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_PIPE_MASTER;
395*f0ea3689SLuigi Rizzo 		nr_ringid = atoi(port + 1);
396*f0ea3689SLuigi Rizzo 		break;
397*f0ea3689SLuigi Rizzo 	case '}':
398*f0ea3689SLuigi Rizzo 		nr_flags = NR_REG_PIPE_SLAVE;
399*f0ea3689SLuigi Rizzo 		nr_ringid = atoi(port + 1);
400*f0ea3689SLuigi Rizzo 		break;
401*f0ea3689SLuigi Rizzo 	}
40266015019SLuigi Rizzo 
403*f0ea3689SLuigi Rizzo 	if (nr_ringid >= NETMAP_RING_MASK) {
404*f0ea3689SLuigi Rizzo 		errmsg = "invalid ringid";
405*f0ea3689SLuigi Rizzo 		goto fail;
406*f0ea3689SLuigi Rizzo 	}
407*f0ea3689SLuigi Rizzo 	/* add the *XPOLL flags */
408*f0ea3689SLuigi Rizzo 	nr_ringid |= new_flags & (NETMAP_NO_TX_POLL | NETMAP_DO_RX_POLL);
409*f0ea3689SLuigi Rizzo 
410*f0ea3689SLuigi Rizzo 	d = (struct nm_desc *)calloc(1, sizeof(*d));
411f9790aebSLuigi Rizzo 	if (d == NULL) {
412*f0ea3689SLuigi Rizzo 		errmsg = "nm_desc alloc failure";
413f9790aebSLuigi Rizzo 		errno = ENOMEM;
414f9790aebSLuigi Rizzo 		return NULL;
415f9790aebSLuigi Rizzo 	}
416f9790aebSLuigi Rizzo 	d->self = d;	/* set this early so nm_close() works */
417f9790aebSLuigi Rizzo 	d->fd = open("/dev/netmap", O_RDWR);
418*f0ea3689SLuigi Rizzo 	if (d->fd < 0) {
419*f0ea3689SLuigi Rizzo 		errmsg = "cannot open /dev/netmap";
420f9790aebSLuigi Rizzo 		goto fail;
421f9790aebSLuigi Rizzo 	}
422*f0ea3689SLuigi Rizzo 
423*f0ea3689SLuigi Rizzo 	if (req)
424*f0ea3689SLuigi Rizzo 		d->req = *req;
425f9790aebSLuigi Rizzo 	d->req.nr_version = NETMAP_API;
426*f0ea3689SLuigi Rizzo 	d->req.nr_ringid &= ~NETMAP_RING_MASK;
427*f0ea3689SLuigi Rizzo 
428*f0ea3689SLuigi Rizzo 	/* these fields are overridden by ifname and flags processing */
429*f0ea3689SLuigi Rizzo 	d->req.nr_ringid |= nr_ringid;
430*f0ea3689SLuigi Rizzo 	d->req.nr_flags = nr_flags;
43166015019SLuigi Rizzo 	memcpy(d->req.nr_name, ifname, namelen);
43266015019SLuigi Rizzo 	d->req.nr_name[namelen] = '\0';
433*f0ea3689SLuigi Rizzo 	/* optionally import info from parent */
434*f0ea3689SLuigi Rizzo 	if (IS_NETMAP_DESC(parent) && new_flags) {
435*f0ea3689SLuigi Rizzo 		if (new_flags & NM_OPEN_ARG1)
436*f0ea3689SLuigi Rizzo 			D("overriding ARG1 %d", parent->req.nr_arg1);
437*f0ea3689SLuigi Rizzo 		d->req.nr_arg1 = new_flags & NM_OPEN_ARG1 ?
438*f0ea3689SLuigi Rizzo 			parent->req.nr_arg1 : 4;
439*f0ea3689SLuigi Rizzo 		if (new_flags & NM_OPEN_ARG2)
440*f0ea3689SLuigi Rizzo 			D("overriding ARG2 %d", parent->req.nr_arg2);
441*f0ea3689SLuigi Rizzo 		d->req.nr_arg2 = new_flags & NM_OPEN_ARG2 ?
442*f0ea3689SLuigi Rizzo 			parent->req.nr_arg2 : 0;
443*f0ea3689SLuigi Rizzo 		if (new_flags & NM_OPEN_ARG3)
444*f0ea3689SLuigi Rizzo 			D("overriding ARG3 %d", parent->req.nr_arg3);
445*f0ea3689SLuigi Rizzo 		d->req.nr_arg3 = new_flags & NM_OPEN_ARG3 ?
446*f0ea3689SLuigi Rizzo 			parent->req.nr_arg3 : 0;
447*f0ea3689SLuigi Rizzo 		if (new_flags & NM_OPEN_RING_CFG) {
448*f0ea3689SLuigi Rizzo 			D("overriding RING_CFG");
449*f0ea3689SLuigi Rizzo 			d->req.nr_tx_slots = parent->req.nr_tx_slots;
450*f0ea3689SLuigi Rizzo 			d->req.nr_rx_slots = parent->req.nr_rx_slots;
451*f0ea3689SLuigi Rizzo 			d->req.nr_tx_rings = parent->req.nr_tx_rings;
452*f0ea3689SLuigi Rizzo 			d->req.nr_rx_rings = parent->req.nr_rx_rings;
453*f0ea3689SLuigi Rizzo 		}
454*f0ea3689SLuigi Rizzo 		if (new_flags & NM_OPEN_IFNAME) {
455*f0ea3689SLuigi Rizzo 			D("overriding ifname %s ringid 0x%x flags 0x%x",
456*f0ea3689SLuigi Rizzo 				parent->req.nr_name, parent->req.nr_ringid,
457*f0ea3689SLuigi Rizzo 				parent->req.nr_flags);
458*f0ea3689SLuigi Rizzo 			memcpy(d->req.nr_name, parent->req.nr_name,
459*f0ea3689SLuigi Rizzo 				sizeof(d->req.nr_name));
460*f0ea3689SLuigi Rizzo 			d->req.nr_ringid = parent->req.nr_ringid;
461*f0ea3689SLuigi Rizzo 			d->req.nr_flags = parent->req.nr_flags;
462*f0ea3689SLuigi Rizzo 		}
463*f0ea3689SLuigi Rizzo 	}
46466015019SLuigi Rizzo 	if (ioctl(d->fd, NIOCREGIF, &d->req)) {
465*f0ea3689SLuigi Rizzo 		errmsg = "NIOCREGIF failed";
466f9790aebSLuigi Rizzo 		goto fail;
46766015019SLuigi Rizzo 	}
468f9790aebSLuigi Rizzo 
469*f0ea3689SLuigi Rizzo 	if (IS_NETMAP_DESC(parent) && parent->mem &&
470*f0ea3689SLuigi Rizzo 	    parent->req.nr_arg2 == d->req.nr_arg2) {
471*f0ea3689SLuigi Rizzo 		/* do not mmap, inherit from parent */
472*f0ea3689SLuigi Rizzo 		d->memsize = parent->memsize;
473*f0ea3689SLuigi Rizzo 		d->mem = parent->mem;
474*f0ea3689SLuigi Rizzo 	} else {
475f9790aebSLuigi Rizzo 		d->memsize = d->req.nr_memsize;
476f9790aebSLuigi Rizzo 		d->mem = mmap(0, d->memsize, PROT_WRITE | PROT_READ, MAP_SHARED,
477f9790aebSLuigi Rizzo 				d->fd, 0);
478*f0ea3689SLuigi Rizzo 		if (d->mem == NULL) {
479*f0ea3689SLuigi Rizzo 			errmsg = "mmap failed";
480f9790aebSLuigi Rizzo 			goto fail;
481*f0ea3689SLuigi Rizzo 		}
482*f0ea3689SLuigi Rizzo 		d->done_mmap = 1;
483*f0ea3689SLuigi Rizzo 	}
484*f0ea3689SLuigi Rizzo 	{
485*f0ea3689SLuigi Rizzo 		struct netmap_if *nifp = NETMAP_IF(d->mem, d->req.nr_offset);
486*f0ea3689SLuigi Rizzo 		struct netmap_ring *r = NETMAP_RXRING(nifp, );
487*f0ea3689SLuigi Rizzo 
488*f0ea3689SLuigi Rizzo 		*(struct netmap_if **)(uintptr_t)&(d->nifp) = nifp;
489*f0ea3689SLuigi Rizzo 		*(struct netmap_ring **)(uintptr_t)&d->some_ring = r;
490*f0ea3689SLuigi Rizzo 		*(void **)(uintptr_t)&d->buf_start = NETMAP_BUF(r, 0);
491*f0ea3689SLuigi Rizzo 		*(void **)(uintptr_t)&d->buf_end =
492*f0ea3689SLuigi Rizzo 			(char *)d->mem + d->memsize;
493*f0ea3689SLuigi Rizzo 	}
494*f0ea3689SLuigi Rizzo 
495*f0ea3689SLuigi Rizzo 	if (nr_flags ==  NR_REG_SW) { /* host stack */
49666015019SLuigi Rizzo 		d->first_tx_ring = d->last_tx_ring = d->req.nr_tx_rings;
49766015019SLuigi Rizzo 		d->first_rx_ring = d->last_rx_ring = d->req.nr_rx_rings;
498*f0ea3689SLuigi Rizzo 	} else if (nr_flags ==  NR_REG_ALL_NIC) { /* only nic */
499*f0ea3689SLuigi Rizzo 		d->first_tx_ring = 0;
500*f0ea3689SLuigi Rizzo 		d->first_rx_ring = 0;
50166015019SLuigi Rizzo 		d->last_tx_ring = d->req.nr_tx_rings - 1;
50266015019SLuigi Rizzo 		d->last_rx_ring = d->req.nr_rx_rings - 1;
503*f0ea3689SLuigi Rizzo 	} else if (nr_flags ==  NR_REG_NIC_SW) {
504*f0ea3689SLuigi Rizzo 		d->first_tx_ring = 0;
505*f0ea3689SLuigi Rizzo 		d->first_rx_ring = 0;
506*f0ea3689SLuigi Rizzo 		d->last_tx_ring = d->req.nr_tx_rings;
507*f0ea3689SLuigi Rizzo 		d->last_rx_ring = d->req.nr_rx_rings;
508*f0ea3689SLuigi Rizzo 	} else if (nr_flags == NR_REG_ONE_NIC) {
509*f0ea3689SLuigi Rizzo 		/* XXX check validity */
510*f0ea3689SLuigi Rizzo 		d->first_tx_ring = d->last_tx_ring =
511*f0ea3689SLuigi Rizzo 		d->first_rx_ring = d->last_rx_ring = nr_ringid;
512*f0ea3689SLuigi Rizzo 	} else { /* pipes */
513*f0ea3689SLuigi Rizzo 		d->first_tx_ring = d->last_tx_ring = 0;
514*f0ea3689SLuigi Rizzo 		d->first_rx_ring = d->last_rx_ring = 0;
515f9790aebSLuigi Rizzo 	}
516*f0ea3689SLuigi Rizzo 
517*f0ea3689SLuigi Rizzo #ifdef DEBUG_NETMAP_USER
518*f0ea3689SLuigi Rizzo     { /* debugging code */
519*f0ea3689SLuigi Rizzo 	int i;
520*f0ea3689SLuigi Rizzo 
521*f0ea3689SLuigi Rizzo 	D("%s tx %d .. %d %d rx %d .. %d %d", ifname,
522*f0ea3689SLuigi Rizzo 		d->first_tx_ring, d->last_tx_ring, d->req.nr_tx_rings,
523*f0ea3689SLuigi Rizzo                 d->first_rx_ring, d->last_rx_ring, d->req.nr_rx_rings);
524*f0ea3689SLuigi Rizzo 	for (i = 0; i <= d->req.nr_tx_rings; i++) {
525*f0ea3689SLuigi Rizzo 		struct netmap_ring *r = NETMAP_TXRING(d->nifp, i);
526*f0ea3689SLuigi Rizzo 		D("TX%d %p h %d c %d t %d", i, r, r->head, r->cur, r->tail);
527*f0ea3689SLuigi Rizzo 	}
528*f0ea3689SLuigi Rizzo 	for (i = 0; i <= d->req.nr_rx_rings; i++) {
529*f0ea3689SLuigi Rizzo 		struct netmap_ring *r = NETMAP_RXRING(d->nifp, i);
530*f0ea3689SLuigi Rizzo 		D("RX%d %p h %d c %d t %d", i, r, r->head, r->cur, r->tail);
531*f0ea3689SLuigi Rizzo 	}
532*f0ea3689SLuigi Rizzo     }
533*f0ea3689SLuigi Rizzo #endif /* debugging */
534*f0ea3689SLuigi Rizzo 
53566015019SLuigi Rizzo 	d->cur_tx_ring = d->first_tx_ring;
53666015019SLuigi Rizzo 	d->cur_rx_ring = d->first_rx_ring;
537f9790aebSLuigi Rizzo 	return d;
538f9790aebSLuigi Rizzo 
539f9790aebSLuigi Rizzo fail:
540f9790aebSLuigi Rizzo 	nm_close(d);
541*f0ea3689SLuigi Rizzo 	if (errmsg)
542*f0ea3689SLuigi Rizzo 		D("%s %s", errmsg, ifname);
543f9790aebSLuigi Rizzo 	errno = EINVAL;
544f9790aebSLuigi Rizzo 	return NULL;
545f9790aebSLuigi Rizzo }
546f9790aebSLuigi Rizzo 
547f9790aebSLuigi Rizzo 
548f9790aebSLuigi Rizzo static int
549*f0ea3689SLuigi Rizzo nm_close(struct nm_desc *d)
550f9790aebSLuigi Rizzo {
55117885a7bSLuigi Rizzo 	/*
55217885a7bSLuigi Rizzo 	 * ugly trick to avoid unused warnings
55317885a7bSLuigi Rizzo 	 */
55417885a7bSLuigi Rizzo 	static void *__xxzt[] __attribute__ ((unused))  =
555b82b2211SLuigi Rizzo 		{ (void *)nm_open, (void *)nm_inject,
556b82b2211SLuigi Rizzo 		  (void *)nm_dispatch, (void *)nm_nextpkt } ;
55717885a7bSLuigi Rizzo 
558f9790aebSLuigi Rizzo 	if (d == NULL || d->self != d)
559f9790aebSLuigi Rizzo 		return EINVAL;
560*f0ea3689SLuigi Rizzo 	if (d->done_mmap && d->mem)
561f9790aebSLuigi Rizzo 		munmap(d->mem, d->memsize);
562f9790aebSLuigi Rizzo 	if (d->fd != -1)
563f9790aebSLuigi Rizzo 		close(d->fd);
564f9790aebSLuigi Rizzo 	bzero(d, sizeof(*d));
565f9790aebSLuigi Rizzo 	free(d);
566f9790aebSLuigi Rizzo 	return 0;
567f9790aebSLuigi Rizzo }
568f9790aebSLuigi Rizzo 
569f9790aebSLuigi Rizzo 
570f9790aebSLuigi Rizzo /*
57117885a7bSLuigi Rizzo  * Same prototype as pcap_inject(), only need to cast.
57217885a7bSLuigi Rizzo  */
57317885a7bSLuigi Rizzo static int
574*f0ea3689SLuigi Rizzo nm_inject(struct nm_desc *d, const void *buf, size_t size)
57517885a7bSLuigi Rizzo {
57666015019SLuigi Rizzo 	u_int c, n = d->last_tx_ring - d->first_tx_ring + 1;
57717885a7bSLuigi Rizzo 
57817885a7bSLuigi Rizzo 	for (c = 0; c < n ; c++) {
57917885a7bSLuigi Rizzo 		/* compute current ring to use */
58017885a7bSLuigi Rizzo 		struct netmap_ring *ring;
58117885a7bSLuigi Rizzo 		uint32_t i, idx;
58266015019SLuigi Rizzo 		uint32_t ri = d->cur_tx_ring + c;
58317885a7bSLuigi Rizzo 
58466015019SLuigi Rizzo 		if (ri > d->last_tx_ring)
58566015019SLuigi Rizzo 			ri = d->first_tx_ring;
58617885a7bSLuigi Rizzo 		ring = NETMAP_TXRING(d->nifp, ri);
58717885a7bSLuigi Rizzo 		if (nm_ring_empty(ring)) {
58817885a7bSLuigi Rizzo 			continue;
58917885a7bSLuigi Rizzo 		}
59017885a7bSLuigi Rizzo 		i = ring->cur;
59117885a7bSLuigi Rizzo 		idx = ring->slot[i].buf_idx;
59217885a7bSLuigi Rizzo 		ring->slot[i].len = size;
593*f0ea3689SLuigi Rizzo 		nm_pkt_copy(buf, NETMAP_BUF(ring, idx), size);
59466015019SLuigi Rizzo 		d->cur_tx_ring = ri;
59517885a7bSLuigi Rizzo 		ring->head = ring->cur = nm_ring_next(ring, i);
59617885a7bSLuigi Rizzo 		return size;
59717885a7bSLuigi Rizzo 	}
59817885a7bSLuigi Rizzo 	return 0; /* fail */
59917885a7bSLuigi Rizzo }
60017885a7bSLuigi Rizzo 
60117885a7bSLuigi Rizzo 
60217885a7bSLuigi Rizzo /*
603f9790aebSLuigi Rizzo  * Same prototype as pcap_dispatch(), only need to cast.
604f9790aebSLuigi Rizzo  */
605f9790aebSLuigi Rizzo static int
606*f0ea3689SLuigi Rizzo nm_dispatch(struct nm_desc *d, int cnt, nm_cb_t cb, u_char *arg)
607f9790aebSLuigi Rizzo {
60866015019SLuigi Rizzo 	int n = d->last_rx_ring - d->first_rx_ring + 1;
60966015019SLuigi Rizzo 	int c, got = 0, ri = d->cur_rx_ring;
610f9790aebSLuigi Rizzo 
611f9790aebSLuigi Rizzo 	if (cnt == 0)
612f9790aebSLuigi Rizzo 		cnt = -1;
613f9790aebSLuigi Rizzo 	/* cnt == -1 means infinite, but rings have a finite amount
614f9790aebSLuigi Rizzo 	 * of buffers and the int is large enough that we never wrap,
615f9790aebSLuigi Rizzo 	 * so we can omit checking for -1
616f9790aebSLuigi Rizzo 	 */
617f9790aebSLuigi Rizzo 	for (c=0; c < n && cnt != got; c++) {
618f9790aebSLuigi Rizzo 		/* compute current ring to use */
619f9790aebSLuigi Rizzo 		struct netmap_ring *ring;
620f9790aebSLuigi Rizzo 
62166015019SLuigi Rizzo 		ri = d->cur_rx_ring + c;
62266015019SLuigi Rizzo 		if (ri > d->last_rx_ring)
62366015019SLuigi Rizzo 			ri = d->first_rx_ring;
624f9790aebSLuigi Rizzo 		ring = NETMAP_RXRING(d->nifp, ri);
62517885a7bSLuigi Rizzo 		for ( ; !nm_ring_empty(ring) && cnt != got; got++) {
626f9790aebSLuigi Rizzo 			u_int i = ring->cur;
627f9790aebSLuigi Rizzo 			u_int idx = ring->slot[i].buf_idx;
628f9790aebSLuigi Rizzo 			u_char *buf = (u_char *)NETMAP_BUF(ring, idx);
62966015019SLuigi Rizzo 
63066015019SLuigi Rizzo 			// __builtin_prefetch(buf);
631f9790aebSLuigi Rizzo 			d->hdr.len = d->hdr.caplen = ring->slot[i].len;
632f9790aebSLuigi Rizzo 			d->hdr.ts = ring->ts;
633f9790aebSLuigi Rizzo 			cb(arg, &d->hdr, buf);
63417885a7bSLuigi Rizzo 			ring->head = ring->cur = nm_ring_next(ring, i);
635f9790aebSLuigi Rizzo 		}
636f9790aebSLuigi Rizzo 	}
63766015019SLuigi Rizzo 	d->cur_rx_ring = ri;
638f9790aebSLuigi Rizzo 	return got;
639f9790aebSLuigi Rizzo }
640f9790aebSLuigi Rizzo 
641f9790aebSLuigi Rizzo static u_char *
642*f0ea3689SLuigi Rizzo nm_nextpkt(struct nm_desc *d, struct nm_pkthdr *hdr)
643f9790aebSLuigi Rizzo {
64466015019SLuigi Rizzo 	int ri = d->cur_rx_ring;
645f9790aebSLuigi Rizzo 
646f9790aebSLuigi Rizzo 	do {
647f9790aebSLuigi Rizzo 		/* compute current ring to use */
648f9790aebSLuigi Rizzo 		struct netmap_ring *ring = NETMAP_RXRING(d->nifp, ri);
64917885a7bSLuigi Rizzo 		if (!nm_ring_empty(ring)) {
650f9790aebSLuigi Rizzo 			u_int i = ring->cur;
651f9790aebSLuigi Rizzo 			u_int idx = ring->slot[i].buf_idx;
652f9790aebSLuigi Rizzo 			u_char *buf = (u_char *)NETMAP_BUF(ring, idx);
65366015019SLuigi Rizzo 
65466015019SLuigi Rizzo 			// __builtin_prefetch(buf);
655f9790aebSLuigi Rizzo 			hdr->ts = ring->ts;
656f9790aebSLuigi Rizzo 			hdr->len = hdr->caplen = ring->slot[i].len;
65717885a7bSLuigi Rizzo 			ring->cur = nm_ring_next(ring, i);
65817885a7bSLuigi Rizzo 			/* we could postpone advancing head if we want
65917885a7bSLuigi Rizzo 			 * to hold the buffer. This can be supported in
66017885a7bSLuigi Rizzo 			 * the future.
66117885a7bSLuigi Rizzo 			 */
66217885a7bSLuigi Rizzo 			ring->head = ring->cur;
66366015019SLuigi Rizzo 			d->cur_rx_ring = ri;
664f9790aebSLuigi Rizzo 			return buf;
665f9790aebSLuigi Rizzo 		}
666f9790aebSLuigi Rizzo 		ri++;
66766015019SLuigi Rizzo 		if (ri > d->last_rx_ring)
66866015019SLuigi Rizzo 			ri = d->first_rx_ring;
66966015019SLuigi Rizzo 	} while (ri != d->cur_rx_ring);
670f9790aebSLuigi Rizzo 	return NULL; /* nothing found */
671f9790aebSLuigi Rizzo }
672f9790aebSLuigi Rizzo 
673f9790aebSLuigi Rizzo #endif /* !HAVE_NETMAP_WITH_LIBS */
674f9790aebSLuigi Rizzo 
675f9790aebSLuigi Rizzo #endif /* NETMAP_WITH_LIBS */
676f9790aebSLuigi Rizzo 
67768b8534bSLuigi Rizzo #endif /* _NET_NETMAP_USER_H_ */
678