xref: /freebsd/sys/dev/netmap/netmap_vale.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (C) 2013-2016 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  *   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 ``AS 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 #if defined(__FreeBSD__)
31 #include <sys/cdefs.h> /* prerequisite */
32 __FBSDID("$FreeBSD$");
33 
34 #include <sys/types.h>
35 #include <sys/errno.h>
36 #include <sys/param.h>	/* defines used in kernel.h */
37 #include <sys/kernel.h>	/* types used in module initialization */
38 #include <sys/conf.h>	/* cdevsw struct, UID, GID */
39 #include <sys/sockio.h>
40 #include <sys/socketvar.h>	/* struct socket */
41 #include <sys/malloc.h>
42 #include <sys/poll.h>
43 #include <sys/rwlock.h>
44 #include <sys/socket.h> /* sockaddrs */
45 #include <sys/selinfo.h>
46 #include <sys/sysctl.h>
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/bpf.h>		/* BIOCIMMEDIATE */
50 #include <machine/bus.h>	/* bus_dmamap_* */
51 #include <sys/endian.h>
52 #include <sys/refcount.h>
53 #include <sys/smp.h>
54 
55 
56 #elif defined(linux)
57 
58 #include "bsd_glue.h"
59 
60 #elif defined(__APPLE__)
61 
62 #warning OSX support is only partial
63 #include "osx_glue.h"
64 
65 #elif defined(_WIN32)
66 #include "win_glue.h"
67 
68 #else
69 
70 #error	Unsupported platform
71 
72 #endif /* unsupported */
73 
74 /*
75  * common headers
76  */
77 
78 #include <net/netmap.h>
79 #include <dev/netmap/netmap_kern.h>
80 #include <dev/netmap/netmap_mem2.h>
81 #include <dev/netmap/netmap_bdg.h>
82 
83 #ifdef WITH_VALE
84 
85 /*
86  * system parameters (most of them in netmap_kern.h)
87  * NM_BDG_NAME		prefix for switch port names, default "vale"
88  * NM_BDG_MAXPORTS	number of ports
89  * NM_BRIDGES		max number of switches in the system.
90  *
91  * Switch ports are named valeX:Y where X is the switch name and Y
92  * is the port. If Y matches a physical interface name, the port is
93  * connected to a physical device.
94  *
95  * Unlike physical interfaces, switch ports use their own memory region
96  * for rings and buffers.
97  * The virtual interfaces use per-queue lock instead of core lock.
98  * In the tx loop, we aggregate traffic in batches to make all operations
99  * faster. The batch size is bridge_batch.
100  */
101 #define NM_BDG_MAXRINGS		16	/* XXX unclear how many (must be a pow of 2). */
102 #define NM_BDG_MAXSLOTS		4096	/* XXX same as above */
103 #define NM_BRIDGE_RINGSIZE	1024	/* in the device */
104 #define NM_BDG_BATCH		1024	/* entries in the forwarding buffer */
105 /* actual size of the tables */
106 #define NM_BDG_BATCH_MAX	(NM_BDG_BATCH + NETMAP_MAX_FRAGS)
107 /* NM_FT_NULL terminates a list of slots in the ft */
108 #define NM_FT_NULL		NM_BDG_BATCH_MAX
109 
110 
111 /*
112  * bridge_batch is set via sysctl to the max batch size to be
113  * used in the bridge. The actual value may be larger as the
114  * last packet in the block may overflow the size.
115  */
116 static int bridge_batch = NM_BDG_BATCH; /* bridge batch size */
117 
118 /* Max number of vale bridges (loader tunable). */
119 unsigned int vale_max_bridges = NM_BRIDGES;
120 
121 SYSBEGIN(vars_vale);
122 SYSCTL_DECL(_dev_netmap);
123 SYSCTL_INT(_dev_netmap, OID_AUTO, bridge_batch, CTLFLAG_RW, &bridge_batch, 0,
124 		"Max batch size to be used in the bridge");
125 SYSCTL_UINT(_dev_netmap, OID_AUTO, max_bridges, CTLFLAG_RDTUN, &vale_max_bridges, 0,
126 		"Max number of vale bridges");
127 SYSEND;
128 
129 static int netmap_vale_vp_create(struct nmreq_header *hdr, struct ifnet *,
130 		struct netmap_mem_d *nmd, struct netmap_vp_adapter **);
131 static int netmap_vale_vp_bdg_attach(const char *, struct netmap_adapter *,
132 		struct nm_bridge *);
133 static int netmap_vale_bwrap_attach(const char *, struct netmap_adapter *);
134 
135 /*
136  * For each output interface, nm_vale_q is used to construct a list.
137  * bq_len is the number of output buffers (we can have coalescing
138  * during the copy).
139  */
140 struct nm_vale_q {
141 	uint16_t bq_head;
142 	uint16_t bq_tail;
143 	uint32_t bq_len;	/* number of buffers */
144 };
145 
146 /* Holds the default callbacks */
147 struct netmap_bdg_ops vale_bdg_ops = {
148 	.lookup = netmap_vale_learning,
149 	.config = NULL,
150 	.dtor = NULL,
151 	.vp_create = netmap_vale_vp_create,
152 	.bwrap_attach = netmap_vale_bwrap_attach,
153 	.name = NM_BDG_NAME,
154 };
155 
156 /*
157  * this is a slightly optimized copy routine which rounds
158  * to multiple of 64 bytes and is often faster than dealing
159  * with other odd sizes. We assume there is enough room
160  * in the source and destination buffers.
161  *
162  * XXX only for multiples of NM_BUF_ALIGN bytes, non overlapped.
163  */
164 
165 static inline void
166 pkt_copy(void *_src, void *_dst, int l)
167 {
168 	uint64_t *src = _src;
169 	uint64_t *dst = _dst;
170 	if (unlikely(l >= 1024)) {
171 		memcpy(dst, src, l);
172 		return;
173 	}
174 	for (; likely(l > 0); l -= NM_BUF_ALIGN) {
175 		/* XXX NM_BUF_ALIGN/sizeof(uint64_t) statements */
176 		*dst++ = *src++;
177 		*dst++ = *src++;
178 		*dst++ = *src++;
179 		*dst++ = *src++;
180 		*dst++ = *src++;
181 		*dst++ = *src++;
182 		*dst++ = *src++;
183 		*dst++ = *src++;
184 	}
185 }
186 
187 
188 /*
189  * Free the forwarding tables for rings attached to switch ports.
190  */
191 static void
192 nm_free_bdgfwd(struct netmap_adapter *na)
193 {
194 	int nrings, i;
195 	struct netmap_kring **kring;
196 
197 	NMG_LOCK_ASSERT();
198 	nrings = na->num_tx_rings;
199 	kring = na->tx_rings;
200 	for (i = 0; i < nrings; i++) {
201 		if (kring[i]->nkr_ft) {
202 			nm_os_free(kring[i]->nkr_ft);
203 			kring[i]->nkr_ft = NULL; /* protect from freeing twice */
204 		}
205 	}
206 }
207 
208 
209 /*
210  * Allocate the forwarding tables for the rings attached to the bridge ports.
211  */
212 static int
213 nm_alloc_bdgfwd(struct netmap_adapter *na)
214 {
215 	int nrings, l, i, num_dstq;
216 	struct netmap_kring **kring;
217 
218 	NMG_LOCK_ASSERT();
219 	/* all port:rings + broadcast */
220 	num_dstq = NM_BDG_MAXPORTS * NM_BDG_MAXRINGS + 1;
221 	l = sizeof(struct nm_bdg_fwd) * NM_BDG_BATCH_MAX;
222 	l += sizeof(struct nm_vale_q) * num_dstq;
223 	l += sizeof(uint16_t) * NM_BDG_BATCH_MAX;
224 
225 	nrings = netmap_real_rings(na, NR_TX);
226 	kring = na->tx_rings;
227 	for (i = 0; i < nrings; i++) {
228 		struct nm_bdg_fwd *ft;
229 		struct nm_vale_q *dstq;
230 		int j;
231 
232 		ft = nm_os_malloc(l);
233 		if (!ft) {
234 			nm_free_bdgfwd(na);
235 			return ENOMEM;
236 		}
237 		dstq = (struct nm_vale_q *)(ft + NM_BDG_BATCH_MAX);
238 		for (j = 0; j < num_dstq; j++) {
239 			dstq[j].bq_head = dstq[j].bq_tail = NM_FT_NULL;
240 			dstq[j].bq_len = 0;
241 		}
242 		kring[i]->nkr_ft = ft;
243 	}
244 	return 0;
245 }
246 
247 /* Allows external modules to create bridges in exclusive mode,
248  * returns an authentication token that the external module will need
249  * to provide during nm_bdg_ctl_{attach, detach}(), netmap_bdg_regops(),
250  * and nm_bdg_update_private_data() operations.
251  * Successfully executed if ret != NULL and *return_status == 0.
252  */
253 void *
254 netmap_vale_create(const char *bdg_name, int *return_status)
255 {
256 	struct nm_bridge *b = NULL;
257 	void *ret = NULL;
258 
259 	NMG_LOCK();
260 	b = nm_find_bridge(bdg_name, 0 /* don't create */, NULL);
261 	if (b) {
262 		*return_status = EEXIST;
263 		goto unlock_bdg_create;
264 	}
265 
266 	b = nm_find_bridge(bdg_name, 1 /* create */, &vale_bdg_ops);
267 	if (!b) {
268 		*return_status = ENOMEM;
269 		goto unlock_bdg_create;
270 	}
271 
272 	b->bdg_flags |= NM_BDG_ACTIVE | NM_BDG_EXCLUSIVE;
273 	ret = nm_bdg_get_auth_token(b);
274 	*return_status = 0;
275 
276 unlock_bdg_create:
277 	NMG_UNLOCK();
278 	return ret;
279 }
280 
281 /* Allows external modules to destroy a bridge created through
282  * netmap_bdg_create(), the bridge must be empty.
283  */
284 int
285 netmap_vale_destroy(const char *bdg_name, void *auth_token)
286 {
287 	struct nm_bridge *b = NULL;
288 	int ret = 0;
289 
290 	NMG_LOCK();
291 	b = nm_find_bridge(bdg_name, 0 /* don't create */, NULL);
292 	if (!b) {
293 		ret = ENXIO;
294 		goto unlock_bdg_free;
295 	}
296 
297 	if (!nm_bdg_valid_auth_token(b, auth_token)) {
298 		ret = EACCES;
299 		goto unlock_bdg_free;
300 	}
301 	if (!(b->bdg_flags & NM_BDG_EXCLUSIVE)) {
302 		ret = EINVAL;
303 		goto unlock_bdg_free;
304 	}
305 
306 	b->bdg_flags &= ~(NM_BDG_EXCLUSIVE | NM_BDG_ACTIVE);
307 	ret = netmap_bdg_free(b);
308 	if (ret) {
309 		b->bdg_flags |= NM_BDG_EXCLUSIVE | NM_BDG_ACTIVE;
310 	}
311 
312 unlock_bdg_free:
313 	NMG_UNLOCK();
314 	return ret;
315 }
316 
317 /* Process NETMAP_REQ_VALE_LIST. */
318 int
319 netmap_vale_list(struct nmreq_header *hdr)
320 {
321 	struct nmreq_vale_list *req =
322 		(struct nmreq_vale_list *)(uintptr_t)hdr->nr_body;
323 	int namelen = strlen(hdr->nr_name);
324 	struct nm_bridge *b, *bridges;
325 	struct netmap_vp_adapter *vpna;
326 	int error = 0, i, j;
327 	u_int num_bridges;
328 
329 	netmap_bns_getbridges(&bridges, &num_bridges);
330 
331 	/* this is used to enumerate bridges and ports */
332 	if (namelen) { /* look up indexes of bridge and port */
333 		if (strncmp(hdr->nr_name, NM_BDG_NAME,
334 					strlen(NM_BDG_NAME))) {
335 			return EINVAL;
336 		}
337 		NMG_LOCK();
338 		b = nm_find_bridge(hdr->nr_name, 0 /* don't create */, NULL);
339 		if (!b) {
340 			NMG_UNLOCK();
341 			return ENOENT;
342 		}
343 
344 		req->nr_bridge_idx = b - bridges; /* bridge index */
345 		req->nr_port_idx = NM_BDG_NOPORT;
346 		for (j = 0; j < b->bdg_active_ports; j++) {
347 			i = b->bdg_port_index[j];
348 			vpna = b->bdg_ports[i];
349 			if (vpna == NULL) {
350 				nm_prerr("This should not happen");
351 				continue;
352 			}
353 			/* the former and the latter identify a
354 			 * virtual port and a NIC, respectively
355 			 */
356 			if (!strcmp(vpna->up.name, hdr->nr_name)) {
357 				req->nr_port_idx = i; /* port index */
358 				break;
359 			}
360 		}
361 		NMG_UNLOCK();
362 	} else {
363 		/* return the first non-empty entry starting from
364 		 * bridge nr_arg1 and port nr_arg2.
365 		 *
366 		 * Users can detect the end of the same bridge by
367 		 * seeing the new and old value of nr_arg1, and can
368 		 * detect the end of all the bridge by error != 0
369 		 */
370 		i = req->nr_bridge_idx;
371 		j = req->nr_port_idx;
372 
373 		NMG_LOCK();
374 		for (error = ENOENT; i < vale_max_bridges; i++) {
375 			b = bridges + i;
376 			for ( ; j < NM_BDG_MAXPORTS; j++) {
377 				if (b->bdg_ports[j] == NULL)
378 					continue;
379 				vpna = b->bdg_ports[j];
380 				/* write back the VALE switch name */
381 				strlcpy(hdr->nr_name, vpna->up.name,
382 					sizeof(hdr->nr_name));
383 				error = 0;
384 				goto out;
385 			}
386 			j = 0; /* following bridges scan from 0 */
387 		}
388 	out:
389 		req->nr_bridge_idx = i;
390 		req->nr_port_idx = j;
391 		NMG_UNLOCK();
392 	}
393 
394 	return error;
395 }
396 
397 
398 /* nm_dtor callback for ephemeral VALE ports */
399 static void
400 netmap_vale_vp_dtor(struct netmap_adapter *na)
401 {
402 	struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter*)na;
403 	struct nm_bridge *b = vpna->na_bdg;
404 
405 	nm_prdis("%s has %d references", na->name, na->na_refcount);
406 
407 	if (b) {
408 		netmap_bdg_detach_common(b, vpna->bdg_port, -1);
409 	}
410 
411 	if (na->ifp != NULL && !nm_iszombie(na)) {
412 		NM_DETACH_NA(na->ifp);
413 		if (vpna->autodelete) {
414 			nm_prdis("releasing %s", na->ifp->if_xname);
415 			NMG_UNLOCK();
416 			nm_os_vi_detach(na->ifp);
417 			NMG_LOCK();
418 		}
419 	}
420 }
421 
422 
423 
424 /* nm_krings_create callback for VALE ports.
425  * Calls the standard netmap_krings_create, then adds leases on rx
426  * rings and bdgfwd on tx rings.
427  */
428 static int
429 netmap_vale_vp_krings_create(struct netmap_adapter *na)
430 {
431 	u_int tailroom;
432 	int error, i;
433 	uint32_t *leases;
434 	u_int nrx = netmap_real_rings(na, NR_RX);
435 
436 	/*
437 	 * Leases are attached to RX rings on vale ports
438 	 */
439 	tailroom = sizeof(uint32_t) * na->num_rx_desc * nrx;
440 
441 	error = netmap_krings_create(na, tailroom);
442 	if (error)
443 		return error;
444 
445 	leases = na->tailroom;
446 
447 	for (i = 0; i < nrx; i++) { /* Receive rings */
448 		na->rx_rings[i]->nkr_leases = leases;
449 		leases += na->num_rx_desc;
450 	}
451 
452 	error = nm_alloc_bdgfwd(na);
453 	if (error) {
454 		netmap_krings_delete(na);
455 		return error;
456 	}
457 
458 	return 0;
459 }
460 
461 
462 /* nm_krings_delete callback for VALE ports. */
463 static void
464 netmap_vale_vp_krings_delete(struct netmap_adapter *na)
465 {
466 	nm_free_bdgfwd(na);
467 	netmap_krings_delete(na);
468 }
469 
470 
471 static int
472 nm_vale_flush(struct nm_bdg_fwd *ft, u_int n,
473 	struct netmap_vp_adapter *na, u_int ring_nr);
474 
475 
476 /*
477  * main dispatch routine for the bridge.
478  * Grab packets from a kring, move them into the ft structure
479  * associated to the tx (input) port. Max one instance per port,
480  * filtered on input (ioctl, poll or XXX).
481  * Returns the next position in the ring.
482  */
483 static int
484 nm_vale_preflush(struct netmap_kring *kring, u_int end)
485 {
486 	struct netmap_vp_adapter *na =
487 		(struct netmap_vp_adapter*)kring->na;
488 	struct netmap_ring *ring = kring->ring;
489 	struct nm_bdg_fwd *ft;
490 	u_int ring_nr = kring->ring_id;
491 	u_int j = kring->nr_hwcur, lim = kring->nkr_num_slots - 1;
492 	u_int ft_i = 0;	/* start from 0 */
493 	u_int frags = 1; /* how many frags ? */
494 	struct nm_bridge *b = na->na_bdg;
495 
496 	/* To protect against modifications to the bridge we acquire a
497 	 * shared lock, waiting if we can sleep (if the source port is
498 	 * attached to a user process) or with a trylock otherwise (NICs).
499 	 */
500 	nm_prdis("wait rlock for %d packets", ((j > end ? lim+1 : 0) + end) - j);
501 	if (na->up.na_flags & NAF_BDG_MAYSLEEP)
502 		BDG_RLOCK(b);
503 	else if (!BDG_RTRYLOCK(b))
504 		return j;
505 	nm_prdis(5, "rlock acquired for %d packets", ((j > end ? lim+1 : 0) + end) - j);
506 	ft = kring->nkr_ft;
507 
508 	for (; likely(j != end); j = nm_next(j, lim)) {
509 		struct netmap_slot *slot = &ring->slot[j];
510 		char *buf;
511 
512 		ft[ft_i].ft_len = slot->len;
513 		ft[ft_i].ft_flags = slot->flags;
514 		ft[ft_i].ft_offset = 0;
515 
516 		nm_prdis("flags is 0x%x", slot->flags);
517 		/* we do not use the buf changed flag, but we still need to reset it */
518 		slot->flags &= ~NS_BUF_CHANGED;
519 
520 		/* this slot goes into a list so initialize the link field */
521 		ft[ft_i].ft_next = NM_FT_NULL;
522 		buf = ft[ft_i].ft_buf = (slot->flags & NS_INDIRECT) ?
523 			(void *)(uintptr_t)slot->ptr : NMB_O(kring, slot);
524 		if (unlikely(buf == NULL ||
525 		     slot->len > NETMAP_BUF_SIZE(&na->up) - nm_get_offset(kring, slot))) {
526 			nm_prlim(5, "NULL %s buffer pointer from %s slot %d len %d",
527 				(slot->flags & NS_INDIRECT) ? "INDIRECT" : "DIRECT",
528 				kring->name, j, ft[ft_i].ft_len);
529 			buf = ft[ft_i].ft_buf = NETMAP_BUF_BASE(&na->up);
530 			ft[ft_i].ft_len = 0;
531 			ft[ft_i].ft_flags = 0;
532 		}
533 		__builtin_prefetch(buf);
534 		++ft_i;
535 		if (slot->flags & NS_MOREFRAG) {
536 			frags++;
537 			continue;
538 		}
539 		if (unlikely(netmap_verbose && frags > 1))
540 			nm_prlim(5, "%d frags at %d", frags, ft_i - frags);
541 		ft[ft_i - frags].ft_frags = frags;
542 		frags = 1;
543 		if (unlikely((int)ft_i >= bridge_batch))
544 			ft_i = nm_vale_flush(ft, ft_i, na, ring_nr);
545 	}
546 	if (frags > 1) {
547 		/* Here ft_i > 0, ft[ft_i-1].flags has NS_MOREFRAG, and we
548 		 * have to fix frags count. */
549 		frags--;
550 		ft[ft_i - 1].ft_flags &= ~NS_MOREFRAG;
551 		ft[ft_i - frags].ft_frags = frags;
552 		nm_prlim(5, "Truncate incomplete fragment at %d (%d frags)", ft_i, frags);
553 	}
554 	if (ft_i)
555 		ft_i = nm_vale_flush(ft, ft_i, na, ring_nr);
556 	BDG_RUNLOCK(b);
557 	return j;
558 }
559 
560 
561 /* ----- FreeBSD if_bridge hash function ------- */
562 
563 /*
564  * The following hash function is adapted from "Hash Functions" by Bob Jenkins
565  * ("Algorithm Alley", Dr. Dobbs Journal, September 1997).
566  *
567  * http://www.burtleburtle.net/bob/hash/spooky.html
568  */
569 #define mix(a, b, c)                                                    \
570 do {                                                                    \
571 	a -= b; a -= c; a ^= (c >> 13);                                 \
572 	b -= c; b -= a; b ^= (a << 8);                                  \
573 	c -= a; c -= b; c ^= (b >> 13);                                 \
574 	a -= b; a -= c; a ^= (c >> 12);                                 \
575 	b -= c; b -= a; b ^= (a << 16);                                 \
576 	c -= a; c -= b; c ^= (b >> 5);                                  \
577 	a -= b; a -= c; a ^= (c >> 3);                                  \
578 	b -= c; b -= a; b ^= (a << 10);                                 \
579 	c -= a; c -= b; c ^= (b >> 15);                                 \
580 } while (/*CONSTCOND*/0)
581 
582 
583 static __inline uint32_t
584 nm_vale_rthash(const uint8_t *addr)
585 {
586 	uint32_t a = 0x9e3779b9, b = 0x9e3779b9, c = 0; // hash key
587 
588 	b += addr[5] << 8;
589 	b += addr[4];
590 	a += addr[3] << 24;
591 	a += addr[2] << 16;
592 	a += addr[1] << 8;
593 	a += addr[0];
594 
595 	mix(a, b, c);
596 #define BRIDGE_RTHASH_MASK	(NM_BDG_HASH-1)
597 	return (c & BRIDGE_RTHASH_MASK);
598 }
599 
600 #undef mix
601 
602 
603 /*
604  * Lookup function for a learning bridge.
605  * Update the hash table with the source address,
606  * and then returns the destination port index, and the
607  * ring in *dst_ring (at the moment, always use ring 0)
608  */
609 uint32_t
610 netmap_vale_learning(struct nm_bdg_fwd *ft, uint8_t *dst_ring,
611 		struct netmap_vp_adapter *na, void *private_data)
612 {
613 	uint8_t *buf = ((uint8_t *)ft->ft_buf) + ft->ft_offset;
614 	u_int buf_len = ft->ft_len - ft->ft_offset;
615 	struct nm_hash_ent *ht = private_data;
616 	uint32_t sh, dh;
617 	u_int dst, mysrc = na->bdg_port;
618 	uint64_t smac, dmac;
619 	uint8_t indbuf[12];
620 
621 	if (buf_len < 14) {
622 		return NM_BDG_NOPORT;
623 	}
624 
625 	if (ft->ft_flags & NS_INDIRECT) {
626 		if (copyin(buf, indbuf, sizeof(indbuf))) {
627 			return NM_BDG_NOPORT;
628 		}
629 		buf = indbuf;
630 	}
631 
632 	dmac = le64toh(*(uint64_t *)(buf)) & 0xffffffffffff;
633 	smac = le64toh(*(uint64_t *)(buf + 4));
634 	smac >>= 16;
635 
636 	/*
637 	 * The hash is somewhat expensive, there might be some
638 	 * worthwhile optimizations here.
639 	 */
640 	if (((buf[6] & 1) == 0) && (na->last_smac != smac)) { /* valid src */
641 		uint8_t *s = buf+6;
642 		sh = nm_vale_rthash(s); /* hash of source */
643 		/* update source port forwarding entry */
644 		na->last_smac = ht[sh].mac = smac;	/* XXX expire ? */
645 		ht[sh].ports = mysrc;
646 		if (netmap_debug & NM_DEBUG_VALE)
647 		    nm_prinf("src %02x:%02x:%02x:%02x:%02x:%02x on port %d",
648 			s[0], s[1], s[2], s[3], s[4], s[5], mysrc);
649 	}
650 	dst = NM_BDG_BROADCAST;
651 	if ((buf[0] & 1) == 0) { /* unicast */
652 		dh = nm_vale_rthash(buf); /* hash of dst */
653 		if (ht[dh].mac == dmac) {	/* found dst */
654 			dst = ht[dh].ports;
655 		}
656 	}
657 	return dst;
658 }
659 
660 
661 /*
662  * Available space in the ring. Only used in VALE code
663  * and only with is_rx = 1
664  */
665 static inline uint32_t
666 nm_kr_space(struct netmap_kring *k, int is_rx)
667 {
668 	int space;
669 
670 	if (is_rx) {
671 		int busy = k->nkr_hwlease - k->nr_hwcur;
672 		if (busy < 0)
673 			busy += k->nkr_num_slots;
674 		space = k->nkr_num_slots - 1 - busy;
675 	} else {
676 		/* XXX never used in this branch */
677 		space = k->nr_hwtail - k->nkr_hwlease;
678 		if (space < 0)
679 			space += k->nkr_num_slots;
680 	}
681 #if 0
682 	// sanity check
683 	if (k->nkr_hwlease >= k->nkr_num_slots ||
684 		k->nr_hwcur >= k->nkr_num_slots ||
685 		k->nr_tail >= k->nkr_num_slots ||
686 		busy < 0 ||
687 		busy >= k->nkr_num_slots) {
688 		nm_prerr("invalid kring, cur %d tail %d lease %d lease_idx %d lim %d",
689 		    k->nr_hwcur, k->nr_hwtail, k->nkr_hwlease,
690 		    k->nkr_lease_idx, k->nkr_num_slots);
691 	}
692 #endif
693 	return space;
694 }
695 
696 
697 
698 
699 /* make a lease on the kring for N positions. return the
700  * lease index
701  * XXX only used in VALE code and with is_rx = 1
702  */
703 static inline uint32_t
704 nm_kr_lease(struct netmap_kring *k, u_int n, int is_rx)
705 {
706 	uint32_t lim = k->nkr_num_slots - 1;
707 	uint32_t lease_idx = k->nkr_lease_idx;
708 
709 	k->nkr_leases[lease_idx] = NR_NOSLOT;
710 	k->nkr_lease_idx = nm_next(lease_idx, lim);
711 
712 #ifdef CONFIG_NETMAP_DEBUG
713 	if (n > nm_kr_space(k, is_rx)) {
714 		nm_prerr("invalid request for %d slots", n);
715 		panic("x");
716 	}
717 #endif /* CONFIG NETMAP_DEBUG */
718 	/* XXX verify that there are n slots */
719 	k->nkr_hwlease += n;
720 	if (k->nkr_hwlease > lim)
721 		k->nkr_hwlease -= lim + 1;
722 
723 #ifdef CONFIG_NETMAP_DEBUG
724 	if (k->nkr_hwlease >= k->nkr_num_slots ||
725 		k->nr_hwcur >= k->nkr_num_slots ||
726 		k->nr_hwtail >= k->nkr_num_slots ||
727 		k->nkr_lease_idx >= k->nkr_num_slots) {
728 		nm_prerr("invalid kring %s, cur %d tail %d lease %d lease_idx %d lim %d",
729 			k->na->name,
730 			k->nr_hwcur, k->nr_hwtail, k->nkr_hwlease,
731 			k->nkr_lease_idx, k->nkr_num_slots);
732 	}
733 #endif /* CONFIG_NETMAP_DEBUG */
734 	return lease_idx;
735 }
736 
737 /*
738  *
739  * This flush routine supports only unicast and broadcast but a large
740  * number of ports, and lets us replace the learn and dispatch functions.
741  */
742 int
743 nm_vale_flush(struct nm_bdg_fwd *ft, u_int n, struct netmap_vp_adapter *na,
744 		u_int ring_nr)
745 {
746 	struct nm_vale_q *dst_ents, *brddst;
747 	uint16_t num_dsts = 0, *dsts;
748 	struct nm_bridge *b = na->na_bdg;
749 	u_int i, me = na->bdg_port;
750 
751 	/*
752 	 * The work area (pointed by ft) is followed by an array of
753 	 * pointers to queues , dst_ents; there are NM_BDG_MAXRINGS
754 	 * queues per port plus one for the broadcast traffic.
755 	 * Then we have an array of destination indexes.
756 	 */
757 	dst_ents = (struct nm_vale_q *)(ft + NM_BDG_BATCH_MAX);
758 	dsts = (uint16_t *)(dst_ents + NM_BDG_MAXPORTS * NM_BDG_MAXRINGS + 1);
759 
760 	/* first pass: find a destination for each packet in the batch */
761 	for (i = 0; likely(i < n); i += ft[i].ft_frags) {
762 		uint8_t dst_ring = ring_nr; /* default, same ring as origin */
763 		uint16_t dst_port, d_i;
764 		struct nm_vale_q *d;
765 		struct nm_bdg_fwd *start_ft = NULL;
766 
767 		nm_prdis("slot %d frags %d", i, ft[i].ft_frags);
768 
769 		if (na->up.virt_hdr_len < ft[i].ft_len) {
770 			ft[i].ft_offset = na->up.virt_hdr_len;
771 			start_ft = &ft[i];
772 		} else if (na->up.virt_hdr_len == ft[i].ft_len && ft[i].ft_flags & NS_MOREFRAG) {
773 			ft[i].ft_offset = ft[i].ft_len;
774 			start_ft = &ft[i+1];
775 		} else {
776 			/* Drop the packet if the virtio-net header is not into the first
777 			 * fragment nor at the very beginning of the second.
778 			 */
779 			continue;
780 		}
781 		dst_port = b->bdg_ops.lookup(start_ft, &dst_ring, na, b->private_data);
782 		if (netmap_verbose > 255)
783 			nm_prlim(5, "slot %d port %d -> %d", i, me, dst_port);
784 		if (dst_port >= NM_BDG_NOPORT)
785 			continue; /* this packet is identified to be dropped */
786 		else if (dst_port == NM_BDG_BROADCAST)
787 			dst_ring = 0; /* broadcasts always go to ring 0 */
788 		else if (unlikely(dst_port == me ||
789 		    !b->bdg_ports[dst_port]))
790 			continue;
791 
792 		/* get a position in the scratch pad */
793 		d_i = dst_port * NM_BDG_MAXRINGS + dst_ring;
794 		d = dst_ents + d_i;
795 
796 		/* append the first fragment to the list */
797 		if (d->bq_head == NM_FT_NULL) { /* new destination */
798 			d->bq_head = d->bq_tail = i;
799 			/* remember this position to be scanned later */
800 			if (dst_port != NM_BDG_BROADCAST)
801 				dsts[num_dsts++] = d_i;
802 		} else {
803 			ft[d->bq_tail].ft_next = i;
804 			d->bq_tail = i;
805 		}
806 		d->bq_len += ft[i].ft_frags;
807 	}
808 
809 	/*
810 	 * Broadcast traffic goes to ring 0 on all destinations.
811 	 * So we need to add these rings to the list of ports to scan.
812 	 */
813 	brddst = dst_ents + NM_BDG_BROADCAST * NM_BDG_MAXRINGS;
814 	if (brddst->bq_head != NM_FT_NULL) {
815 		u_int j;
816 		for (j = 0; likely(j < b->bdg_active_ports); j++) {
817 			uint16_t d_i;
818 			i = b->bdg_port_index[j];
819 			if (unlikely(i == me))
820 				continue;
821 			d_i = i * NM_BDG_MAXRINGS;
822 			if (dst_ents[d_i].bq_head == NM_FT_NULL)
823 				dsts[num_dsts++] = d_i;
824 		}
825 	}
826 
827 	nm_prdis(5, "pass 1 done %d pkts %d dsts", n, num_dsts);
828 	/* second pass: scan destinations */
829 	for (i = 0; i < num_dsts; i++) {
830 		struct netmap_vp_adapter *dst_na;
831 		struct netmap_kring *kring;
832 		struct netmap_ring *ring;
833 		u_int dst_nr, lim, j, d_i, next, brd_next;
834 		u_int needed, howmany;
835 		int retry = netmap_txsync_retry;
836 		struct nm_vale_q *d;
837 		uint32_t my_start = 0, lease_idx = 0;
838 		int nrings;
839 		int virt_hdr_mismatch = 0;
840 
841 		d_i = dsts[i];
842 		nm_prdis("second pass %d port %d", i, d_i);
843 		d = dst_ents + d_i;
844 		// XXX fix the division
845 		dst_na = b->bdg_ports[d_i/NM_BDG_MAXRINGS];
846 		/* protect from the lookup function returning an inactive
847 		 * destination port
848 		 */
849 		if (unlikely(dst_na == NULL))
850 			goto cleanup;
851 		if (dst_na->up.na_flags & NAF_SW_ONLY)
852 			goto cleanup;
853 		/*
854 		 * The interface may be in !netmap mode in two cases:
855 		 * - when na is attached but not activated yet;
856 		 * - when na is being deactivated but is still attached.
857 		 */
858 		if (unlikely(!nm_netmap_on(&dst_na->up))) {
859 			nm_prdis("not in netmap mode!");
860 			goto cleanup;
861 		}
862 
863 		/* there is at least one either unicast or broadcast packet */
864 		brd_next = brddst->bq_head;
865 		next = d->bq_head;
866 		/* we need to reserve this many slots. If fewer are
867 		 * available, some packets will be dropped.
868 		 * Packets may have multiple fragments, so
869 		 * there is a chance that we may not use all of the slots
870 		 * we have claimed, so we will need to handle the leftover
871 		 * ones when we regain the lock.
872 		 */
873 		needed = d->bq_len + brddst->bq_len;
874 
875 		if (unlikely(dst_na->up.virt_hdr_len != na->up.virt_hdr_len)) {
876 			if (netmap_verbose) {
877 				nm_prlim(3, "virt_hdr_mismatch, src %d dst %d", na->up.virt_hdr_len,
878 						dst_na->up.virt_hdr_len);
879 			}
880 			/* There is a virtio-net header/offloadings mismatch between
881 			 * source and destination. The slower mismatch datapath will
882 			 * be used to cope with all the mismatches.
883 			 */
884 			virt_hdr_mismatch = 1;
885 			if (dst_na->mfs < na->mfs) {
886 				/* We may need to do segmentation offloadings, and so
887 				 * we may need a number of destination slots greater
888 				 * than the number of input slots ('needed').
889 				 * We look for the smallest integer 'x' which satisfies:
890 				 *	needed * na->mfs + x * H <= x * na->mfs
891 				 * where 'H' is the length of the longest header that may
892 				 * be replicated in the segmentation process (e.g. for
893 				 * TCPv4 we must account for ethernet header, IP header
894 				 * and TCPv4 header).
895 				 */
896 				KASSERT(dst_na->mfs > 0, ("vpna->mfs is 0"));
897 				needed = (needed * na->mfs) /
898 						(dst_na->mfs - WORST_CASE_GSO_HEADER) + 1;
899 				nm_prdis(3, "srcmtu=%u, dstmtu=%u, x=%u", na->mfs, dst_na->mfs, needed);
900 			}
901 		}
902 
903 		nm_prdis(5, "pass 2 dst %d is %x %s",
904 			i, d_i, nm_is_bwrap(&dst_na->up) ? "nic/host" : "virtual");
905 		dst_nr = d_i & (NM_BDG_MAXRINGS-1);
906 		nrings = dst_na->up.num_rx_rings;
907 		if (dst_nr >= nrings)
908 			dst_nr = dst_nr % nrings;
909 		kring = dst_na->up.rx_rings[dst_nr];
910 		ring = kring->ring;
911 		/* the destination ring may have not been opened for RX */
912 		if (unlikely(ring == NULL || kring->nr_mode != NKR_NETMAP_ON))
913 			goto cleanup;
914 		lim = kring->nkr_num_slots - 1;
915 
916 retry:
917 
918 		if (dst_na->retry && retry) {
919 			/* try to get some free slot from the previous run */
920 			kring->nm_notify(kring, NAF_FORCE_RECLAIM);
921 			/* actually useful only for bwraps, since there
922 			 * the notify will trigger a txsync on the hwna. VALE ports
923 			 * have dst_na->retry == 0
924 			 */
925 		}
926 		/* reserve the buffers in the queue and an entry
927 		 * to report completion, and drop lock.
928 		 * XXX this might become a helper function.
929 		 */
930 		mtx_lock(&kring->q_lock);
931 		if (kring->nkr_stopped) {
932 			mtx_unlock(&kring->q_lock);
933 			goto cleanup;
934 		}
935 		my_start = j = kring->nkr_hwlease;
936 		howmany = nm_kr_space(kring, 1);
937 		if (needed < howmany)
938 			howmany = needed;
939 		lease_idx = nm_kr_lease(kring, howmany, 1);
940 		mtx_unlock(&kring->q_lock);
941 
942 		/* only retry if we need more than available slots */
943 		if (retry && needed <= howmany)
944 			retry = 0;
945 
946 		/* copy to the destination queue */
947 		while (howmany > 0) {
948 			struct netmap_slot *slot;
949 			struct nm_bdg_fwd *ft_p, *ft_end;
950 			u_int cnt;
951 
952 			/* find the queue from which we pick next packet.
953 			 * NM_FT_NULL is always higher than valid indexes
954 			 * so we never dereference it if the other list
955 			 * has packets (and if both are empty we never
956 			 * get here).
957 			 */
958 			if (next < brd_next) {
959 				ft_p = ft + next;
960 				next = ft_p->ft_next;
961 			} else { /* insert broadcast */
962 				ft_p = ft + brd_next;
963 				brd_next = ft_p->ft_next;
964 			}
965 			cnt = ft_p->ft_frags; // cnt > 0
966 			if (unlikely(cnt > howmany))
967 			    break; /* no more space */
968 			if (netmap_verbose && cnt > 1)
969 				nm_prlim(5, "rx %d frags to %d", cnt, j);
970 			ft_end = ft_p + cnt;
971 			if (unlikely(virt_hdr_mismatch)) {
972 				bdg_mismatch_datapath(na, dst_na, ft_p, ring, &j, lim, &howmany);
973 			} else {
974 				howmany -= cnt;
975 				do {
976 					char *dst, *src = ft_p->ft_buf;
977 					size_t copy_len = ft_p->ft_len, dst_len = copy_len;
978 					uintptr_t src_cb;
979 					uint64_t dstoff, dstoff_cb;
980 					int src_co, dst_co;
981 					const uintptr_t mask = NM_BUF_ALIGN - 1;
982 
983 					slot = &ring->slot[j];
984 					dst = NMB(&dst_na->up, slot);
985 					dstoff = nm_get_offset(kring, slot);
986 					dstoff_cb = dstoff & ~mask;
987 					src_cb = ((uintptr_t)src) & ~mask;
988 					src_co = ((uintptr_t)src) & mask;
989 					dst_co = ((uintptr_t)(dst + dstoff)) & mask;
990 					if (dst_co < src_co) {
991 						dstoff_cb += NM_BUF_ALIGN;
992 					}
993 					dstoff = dstoff_cb + src_co;
994 					copy_len += src_co;
995 
996 					nm_prdis("send [%d] %d(%d) bytes at %s:%d",
997 							i, (int)copy_len, (int)dst_len,
998 							NM_IFPNAME(dst_ifp), j);
999 
1000 					if (unlikely(dstoff > NETMAP_BUF_SIZE(&dst_na->up) ||
1001 				                     dst_len > NETMAP_BUF_SIZE(&dst_na->up) - dstoff)) {
1002 						nm_prlim(5, "dropping packet/fragment of len %zu, dest offset %llu",
1003 								dst_len, (unsigned long long)dstoff);
1004 						copy_len = dst_len = 0;
1005 						dstoff = nm_get_offset(kring, slot);
1006 					}
1007 
1008 					if (ft_p->ft_flags & NS_INDIRECT) {
1009 						if (copyin(src, dst, copy_len)) {
1010 							// invalid user pointer, pretend len is 0
1011 							dst_len = 0;
1012 						}
1013 					} else {
1014 						//memcpy(dst, src, copy_len);
1015 						pkt_copy((char *)src_cb, dst + dstoff_cb, (int)copy_len);
1016 					}
1017 					slot->len = dst_len;
1018 					slot->flags = (cnt << 8)| NS_MOREFRAG;
1019 					nm_write_offset(kring, slot, dstoff);
1020 					j = nm_next(j, lim);
1021 					needed--;
1022 					ft_p++;
1023 				} while (ft_p != ft_end);
1024 				slot->flags = (cnt << 8); /* clear flag on last entry */
1025 			}
1026 			/* are we done ? */
1027 			if (next == NM_FT_NULL && brd_next == NM_FT_NULL)
1028 				break;
1029 		}
1030 		{
1031 		    /* current position */
1032 		    uint32_t *p = kring->nkr_leases; /* shorthand */
1033 		    uint32_t update_pos;
1034 		    int still_locked = 1;
1035 
1036 		    mtx_lock(&kring->q_lock);
1037 		    if (unlikely(howmany > 0)) {
1038 			/* not used all bufs. If i am the last one
1039 			 * i can recover the slots, otherwise must
1040 			 * fill them with 0 to mark empty packets.
1041 			 */
1042 			nm_prdis("leftover %d bufs", howmany);
1043 			if (nm_next(lease_idx, lim) == kring->nkr_lease_idx) {
1044 			    /* yes i am the last one */
1045 			    nm_prdis("roll back nkr_hwlease to %d", j);
1046 			    kring->nkr_hwlease = j;
1047 			} else {
1048 			    while (howmany-- > 0) {
1049 				ring->slot[j].len = 0;
1050 				ring->slot[j].flags = 0;
1051 				j = nm_next(j, lim);
1052 			    }
1053 			}
1054 		    }
1055 		    p[lease_idx] = j; /* report I am done */
1056 
1057 		    update_pos = kring->nr_hwtail;
1058 
1059 		    if (my_start == update_pos) {
1060 			/* all slots before my_start have been reported,
1061 			 * so scan subsequent leases to see if other ranges
1062 			 * have been completed, and to a selwakeup or txsync.
1063 		         */
1064 			while (lease_idx != kring->nkr_lease_idx &&
1065 				p[lease_idx] != NR_NOSLOT) {
1066 			    j = p[lease_idx];
1067 			    p[lease_idx] = NR_NOSLOT;
1068 			    lease_idx = nm_next(lease_idx, lim);
1069 			}
1070 			/* j is the new 'write' position. j != my_start
1071 			 * means there are new buffers to report
1072 			 */
1073 			if (likely(j != my_start)) {
1074 				kring->nr_hwtail = j;
1075 				still_locked = 0;
1076 				mtx_unlock(&kring->q_lock);
1077 				kring->nm_notify(kring, 0);
1078 				/* this is netmap_notify for VALE ports and
1079 				 * netmap_bwrap_notify for bwrap. The latter will
1080 				 * trigger a txsync on the underlying hwna
1081 				 */
1082 				if (dst_na->retry && retry--) {
1083 					/* XXX this is going to call nm_notify again.
1084 					 * Only useful for bwrap in virtual machines
1085 					 */
1086 					goto retry;
1087 				}
1088 			}
1089 		    }
1090 		    if (still_locked)
1091 			mtx_unlock(&kring->q_lock);
1092 		}
1093 cleanup:
1094 		d->bq_head = d->bq_tail = NM_FT_NULL; /* cleanup */
1095 		d->bq_len = 0;
1096 	}
1097 	brddst->bq_head = brddst->bq_tail = NM_FT_NULL; /* cleanup */
1098 	brddst->bq_len = 0;
1099 	return 0;
1100 }
1101 
1102 /* nm_txsync callback for VALE ports */
1103 static int
1104 netmap_vale_vp_txsync(struct netmap_kring *kring, int flags)
1105 {
1106 	struct netmap_vp_adapter *na =
1107 		(struct netmap_vp_adapter *)kring->na;
1108 	u_int done;
1109 	u_int const lim = kring->nkr_num_slots - 1;
1110 	u_int const head = kring->rhead;
1111 
1112 	if (bridge_batch <= 0) { /* testing only */
1113 		done = head; // used all
1114 		goto done;
1115 	}
1116 	if (!na->na_bdg) {
1117 		done = head;
1118 		goto done;
1119 	}
1120 	if (bridge_batch > NM_BDG_BATCH)
1121 		bridge_batch = NM_BDG_BATCH;
1122 
1123 	done = nm_vale_preflush(kring, head);
1124 done:
1125 	if (done != head)
1126 		nm_prerr("early break at %d/ %d, tail %d", done, head, kring->nr_hwtail);
1127 	/*
1128 	 * packets between 'done' and 'cur' are left unsent.
1129 	 */
1130 	kring->nr_hwcur = done;
1131 	kring->nr_hwtail = nm_prev(done, lim);
1132 	if (netmap_debug & NM_DEBUG_TXSYNC)
1133 		nm_prinf("%s ring %d flags %d", na->up.name, kring->ring_id, flags);
1134 	return 0;
1135 }
1136 
1137 
1138 /* create a netmap_vp_adapter that describes a VALE port.
1139  * Only persistent VALE ports have a non-null ifp.
1140  */
1141 static int
1142 netmap_vale_vp_create(struct nmreq_header *hdr, struct ifnet *ifp,
1143 		struct netmap_mem_d *nmd, struct netmap_vp_adapter **ret)
1144 {
1145 	struct nmreq_register *req = (struct nmreq_register *)(uintptr_t)hdr->nr_body;
1146 	struct netmap_vp_adapter *vpna;
1147 	struct netmap_adapter *na;
1148 	int error = 0;
1149 	u_int npipes = 0;
1150 	u_int extrabufs = 0;
1151 
1152 	if (hdr->nr_reqtype != NETMAP_REQ_REGISTER) {
1153 		return EINVAL;
1154 	}
1155 
1156 	vpna = nm_os_malloc(sizeof(*vpna));
1157 	if (vpna == NULL)
1158 		return ENOMEM;
1159 
1160  	na = &vpna->up;
1161 
1162 	na->ifp = ifp;
1163 	strlcpy(na->name, hdr->nr_name, sizeof(na->name));
1164 
1165 	/* bound checking */
1166 	na->num_tx_rings = req->nr_tx_rings;
1167 	nm_bound_var(&na->num_tx_rings, 1, 1, NM_BDG_MAXRINGS, NULL);
1168 	req->nr_tx_rings = na->num_tx_rings; /* write back */
1169 	na->num_rx_rings = req->nr_rx_rings;
1170 	nm_bound_var(&na->num_rx_rings, 1, 1, NM_BDG_MAXRINGS, NULL);
1171 	req->nr_rx_rings = na->num_rx_rings; /* write back */
1172 	nm_bound_var(&req->nr_tx_slots, NM_BRIDGE_RINGSIZE,
1173 			1, NM_BDG_MAXSLOTS, NULL);
1174 	na->num_tx_desc = req->nr_tx_slots;
1175 	nm_bound_var(&req->nr_rx_slots, NM_BRIDGE_RINGSIZE,
1176 			1, NM_BDG_MAXSLOTS, NULL);
1177 	/* validate number of pipes. We want at least 1,
1178 	 * but probably can do with some more.
1179 	 * So let's use 2 as default (when 0 is supplied)
1180 	 */
1181 	nm_bound_var(&npipes, 2, 1, NM_MAXPIPES, NULL);
1182 	/* validate extra bufs */
1183 	extrabufs = req->nr_extra_bufs;
1184 	nm_bound_var(&extrabufs, 0, 0,
1185 			128*NM_BDG_MAXSLOTS, NULL);
1186 	req->nr_extra_bufs = extrabufs; /* write back */
1187 	na->num_rx_desc = req->nr_rx_slots;
1188 	/* Set the mfs to a default value, as it is needed on the VALE
1189 	 * mismatch datapath. XXX We should set it according to the MTU
1190 	 * known to the kernel. */
1191 	vpna->mfs = NM_BDG_MFS_DEFAULT;
1192 	vpna->last_smac = ~0llu;
1193 	/*if (vpna->mfs > netmap_buf_size)  TODO netmap_buf_size is zero??
1194 		vpna->mfs = netmap_buf_size; */
1195 	if (netmap_verbose)
1196 		nm_prinf("max frame size %u", vpna->mfs);
1197 
1198 	na->na_flags |= (NAF_BDG_MAYSLEEP | NAF_OFFSETS);
1199 	/* persistent VALE ports look like hw devices
1200 	 * with a native netmap adapter
1201 	 */
1202 	if (ifp)
1203 		na->na_flags |= NAF_NATIVE;
1204 	na->nm_txsync = netmap_vale_vp_txsync;
1205 	na->nm_rxsync = netmap_vp_rxsync; /* use the one provided by bdg */
1206 	na->nm_register = netmap_vp_reg;  /* use the one provided by bdg */
1207 	na->nm_krings_create = netmap_vale_vp_krings_create;
1208 	na->nm_krings_delete = netmap_vale_vp_krings_delete;
1209 	na->nm_dtor = netmap_vale_vp_dtor;
1210 	nm_prdis("nr_mem_id %d", req->nr_mem_id);
1211 	na->nm_mem = nmd ?
1212 		netmap_mem_get(nmd):
1213 		netmap_mem_private_new(
1214 			na->num_tx_rings, na->num_tx_desc,
1215 			na->num_rx_rings, na->num_rx_desc,
1216 			req->nr_extra_bufs, npipes, &error);
1217 	if (na->nm_mem == NULL)
1218 		goto err;
1219 	na->nm_bdg_attach = netmap_vale_vp_bdg_attach;
1220 	/* other nmd fields are set in the common routine */
1221 	error = netmap_attach_common(na);
1222 	if (error)
1223 		goto err;
1224 	*ret = vpna;
1225 	return 0;
1226 
1227 err:
1228 	if (na->nm_mem != NULL)
1229 		netmap_mem_put(na->nm_mem);
1230 	nm_os_free(vpna);
1231 	return error;
1232 }
1233 
1234 /* nm_bdg_attach callback for VALE ports
1235  * The na_vp port is this same netmap_adapter. There is no host port.
1236  */
1237 static int
1238 netmap_vale_vp_bdg_attach(const char *name, struct netmap_adapter *na,
1239 		struct nm_bridge *b)
1240 {
1241 	struct netmap_vp_adapter *vpna = (struct netmap_vp_adapter *)na;
1242 
1243 	if ((b->bdg_flags & NM_BDG_NEED_BWRAP) || vpna->na_bdg) {
1244 		return NM_NEED_BWRAP;
1245 	}
1246 	na->na_vp = vpna;
1247 	strlcpy(na->name, name, sizeof(na->name));
1248 	na->na_hostvp = NULL;
1249 	return 0;
1250 }
1251 
1252 static int
1253 netmap_vale_bwrap_krings_create(struct netmap_adapter *na)
1254 {
1255 	int error;
1256 
1257 	/* impersonate a netmap_vp_adapter */
1258 	error = netmap_vale_vp_krings_create(na);
1259 	if (error)
1260 		return error;
1261 	error = netmap_bwrap_krings_create_common(na);
1262 	if (error) {
1263 		netmap_vale_vp_krings_delete(na);
1264 	}
1265 	return error;
1266 }
1267 
1268 static void
1269 netmap_vale_bwrap_krings_delete(struct netmap_adapter *na)
1270 {
1271 	netmap_bwrap_krings_delete_common(na);
1272 	netmap_vale_vp_krings_delete(na);
1273 }
1274 
1275 static int
1276 netmap_vale_bwrap_attach(const char *nr_name, struct netmap_adapter *hwna)
1277 {
1278 	struct netmap_bwrap_adapter *bna;
1279 	struct netmap_adapter *na = NULL;
1280 	struct netmap_adapter *hostna = NULL;
1281 	int error;
1282 
1283 	bna = nm_os_malloc(sizeof(*bna));
1284 	if (bna == NULL) {
1285 		return ENOMEM;
1286 	}
1287 	na = &bna->up.up;
1288 	strlcpy(na->name, nr_name, sizeof(na->name));
1289 	na->nm_register = netmap_bwrap_reg;
1290 	na->nm_txsync = netmap_vale_vp_txsync;
1291 	// na->nm_rxsync = netmap_bwrap_rxsync;
1292 	na->nm_krings_create = netmap_vale_bwrap_krings_create;
1293 	na->nm_krings_delete = netmap_vale_bwrap_krings_delete;
1294 	na->nm_notify = netmap_bwrap_notify;
1295 	bna->nm_intr_notify = netmap_bwrap_intr_notify;
1296 	bna->up.retry = 1; /* XXX maybe this should depend on the hwna */
1297 	/* Set the mfs, needed on the VALE mismatch datapath. */
1298 	bna->up.mfs = NM_BDG_MFS_DEFAULT;
1299 
1300 	if (hwna->na_flags & NAF_HOST_RINGS) {
1301 		hostna = &bna->host.up;
1302 		hostna->nm_notify = netmap_bwrap_notify;
1303 		bna->host.mfs = NM_BDG_MFS_DEFAULT;
1304 	}
1305 
1306 	error = netmap_bwrap_attach_common(na, hwna);
1307 	if (error) {
1308 		nm_os_free(bna);
1309 	}
1310 	return error;
1311 }
1312 
1313 int
1314 netmap_get_vale_na(struct nmreq_header *hdr, struct netmap_adapter **na,
1315 		struct netmap_mem_d *nmd, int create)
1316 {
1317 	return netmap_get_bdg_na(hdr, na, nmd, create, &vale_bdg_ops);
1318 }
1319 
1320 
1321 /* creates a persistent VALE port */
1322 int
1323 nm_vi_create(struct nmreq_header *hdr)
1324 {
1325 	struct nmreq_vale_newif *req =
1326 		(struct nmreq_vale_newif *)(uintptr_t)hdr->nr_body;
1327 	int error = 0;
1328 	/* Build a nmreq_register out of the nmreq_vale_newif,
1329 	 * so that we can call netmap_get_bdg_na(). */
1330 	struct nmreq_register regreq;
1331 	bzero(&regreq, sizeof(regreq));
1332 	regreq.nr_tx_slots = req->nr_tx_slots;
1333 	regreq.nr_rx_slots = req->nr_rx_slots;
1334 	regreq.nr_tx_rings = req->nr_tx_rings;
1335 	regreq.nr_rx_rings = req->nr_rx_rings;
1336 	regreq.nr_mem_id = req->nr_mem_id;
1337 	hdr->nr_reqtype = NETMAP_REQ_REGISTER;
1338 	hdr->nr_body = (uintptr_t)&regreq;
1339 	error = netmap_vi_create(hdr, 0 /* no autodelete */);
1340 	hdr->nr_reqtype = NETMAP_REQ_VALE_NEWIF;
1341 	hdr->nr_body = (uintptr_t)req;
1342 	/* Write back to the original struct. */
1343 	req->nr_tx_slots = regreq.nr_tx_slots;
1344 	req->nr_rx_slots = regreq.nr_rx_slots;
1345 	req->nr_tx_rings = regreq.nr_tx_rings;
1346 	req->nr_rx_rings = regreq.nr_rx_rings;
1347 	req->nr_mem_id = regreq.nr_mem_id;
1348 	return error;
1349 }
1350 
1351 /* remove a persistent VALE port from the system */
1352 int
1353 nm_vi_destroy(const char *name)
1354 {
1355 	struct ifnet *ifp;
1356 	struct netmap_vp_adapter *vpna;
1357 	int error;
1358 
1359 	ifp = ifunit_ref(name);
1360 	if (!ifp)
1361 		return ENXIO;
1362 	NMG_LOCK();
1363 	/* make sure this is actually a VALE port */
1364 	if (!NM_NA_VALID(ifp) || NA(ifp)->nm_register != netmap_vp_reg) {
1365 		error = EINVAL;
1366 		goto err;
1367 	}
1368 
1369 	vpna = (struct netmap_vp_adapter *)NA(ifp);
1370 
1371 	/* we can only destroy ports that were created via NETMAP_BDG_NEWIF */
1372 	if (vpna->autodelete) {
1373 		error = EINVAL;
1374 		goto err;
1375 	}
1376 
1377 	/* also make sure that nobody is using the interface */
1378 	if (NETMAP_OWNED_BY_ANY(&vpna->up) ||
1379 	    vpna->up.na_refcount > 1 /* any ref besides the one in nm_vi_create()? */) {
1380 		error = EBUSY;
1381 		goto err;
1382 	}
1383 
1384 	NMG_UNLOCK();
1385 
1386 	if (netmap_verbose)
1387 		nm_prinf("destroying a persistent vale interface %s", ifp->if_xname);
1388 	/* Linux requires all the references are released
1389 	 * before unregister
1390 	 */
1391 	netmap_detach(ifp);
1392 	if_rele(ifp);
1393 	nm_os_vi_detach(ifp);
1394 	return 0;
1395 
1396 err:
1397 	NMG_UNLOCK();
1398 	if_rele(ifp);
1399 	return error;
1400 }
1401 
1402 static int
1403 nm_update_info(struct nmreq_register *req, struct netmap_adapter *na)
1404 {
1405 	req->nr_rx_rings = na->num_rx_rings;
1406 	req->nr_tx_rings = na->num_tx_rings;
1407 	req->nr_rx_slots = na->num_rx_desc;
1408 	req->nr_tx_slots = na->num_tx_desc;
1409 	return netmap_mem_get_info(na->nm_mem, &req->nr_memsize, NULL,
1410 					&req->nr_mem_id);
1411 }
1412 
1413 
1414 /*
1415  * Create a virtual interface registered to the system.
1416  * The interface will be attached to a bridge later.
1417  */
1418 int
1419 netmap_vi_create(struct nmreq_header *hdr, int autodelete)
1420 {
1421 	struct nmreq_register *req = (struct nmreq_register *)(uintptr_t)hdr->nr_body;
1422 	struct ifnet *ifp;
1423 	struct netmap_vp_adapter *vpna;
1424 	struct netmap_mem_d *nmd = NULL;
1425 	int error;
1426 
1427 	if (hdr->nr_reqtype != NETMAP_REQ_REGISTER) {
1428 		return EINVAL;
1429 	}
1430 
1431 	/* don't include VALE prefix */
1432 	if (!strncmp(hdr->nr_name, NM_BDG_NAME, strlen(NM_BDG_NAME)))
1433 		return EINVAL;
1434 	if (strlen(hdr->nr_name) >= IFNAMSIZ) {
1435 		return EINVAL;
1436 	}
1437 	ifp = ifunit_ref(hdr->nr_name);
1438 	if (ifp) { /* already exist, cannot create new one */
1439 		error = EEXIST;
1440 		NMG_LOCK();
1441 		if (NM_NA_VALID(ifp)) {
1442 			int update_err = nm_update_info(req, NA(ifp));
1443 			if (update_err)
1444 				error = update_err;
1445 		}
1446 		NMG_UNLOCK();
1447 		if_rele(ifp);
1448 		return error;
1449 	}
1450 	error = nm_os_vi_persist(hdr->nr_name, &ifp);
1451 	if (error)
1452 		return error;
1453 
1454 	NMG_LOCK();
1455 	if (req->nr_mem_id) {
1456 		nmd = netmap_mem_find(req->nr_mem_id);
1457 		if (nmd == NULL) {
1458 			error = EINVAL;
1459 			goto err_1;
1460 		}
1461 	}
1462 	/* netmap_vp_create creates a struct netmap_vp_adapter */
1463 	error = netmap_vale_vp_create(hdr, ifp, nmd, &vpna);
1464 	if (error) {
1465 		if (netmap_debug & NM_DEBUG_VALE)
1466 			nm_prerr("error %d", error);
1467 		goto err_1;
1468 	}
1469 	/* persist-specific routines */
1470 	vpna->up.nm_bdg_ctl = netmap_vp_bdg_ctl;
1471 	if (!autodelete) {
1472 		netmap_adapter_get(&vpna->up);
1473 	} else {
1474 		vpna->autodelete = 1;
1475 	}
1476 	NM_ATTACH_NA(ifp, &vpna->up);
1477 	/* return the updated info */
1478 	error = nm_update_info(req, &vpna->up);
1479 	if (error) {
1480 		goto err_2;
1481 	}
1482 	nm_prdis("returning nr_mem_id %d", req->nr_mem_id);
1483 	if (nmd)
1484 		netmap_mem_put(nmd);
1485 	NMG_UNLOCK();
1486 	nm_prdis("created %s", ifp->if_xname);
1487 	return 0;
1488 
1489 err_2:
1490 	netmap_detach(ifp);
1491 err_1:
1492 	if (nmd)
1493 		netmap_mem_put(nmd);
1494 	NMG_UNLOCK();
1495 	nm_os_vi_detach(ifp);
1496 
1497 	return error;
1498 }
1499 
1500 #endif /* WITH_VALE */
1501