xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision d37eb51047221dc3322b34db1038ff3aa533883f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2004-2006 Kip Macy
5  * Copyright (c) 2015 Wei Liu <wei.liu2@citrix.com>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
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 #include <sys/cdefs.h>
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 
34 #include <sys/param.h>
35 #include <sys/sockio.h>
36 #include <sys/limits.h>
37 #include <sys/mbuf.h>
38 #include <sys/malloc.h>
39 #include <sys/module.h>
40 #include <sys/kernel.h>
41 #include <sys/socket.h>
42 #include <sys/sysctl.h>
43 #include <sys/taskqueue.h>
44 
45 #include <net/if.h>
46 #include <net/if_var.h>
47 #include <net/if_arp.h>
48 #include <net/ethernet.h>
49 #include <net/if_media.h>
50 #include <net/bpf.h>
51 #include <net/if_types.h>
52 
53 #include <netinet/in.h>
54 #include <netinet/ip.h>
55 #include <netinet/if_ether.h>
56 #include <netinet/tcp.h>
57 #include <netinet/tcp_lro.h>
58 
59 #include <vm/vm.h>
60 #include <vm/pmap.h>
61 
62 #include <sys/bus.h>
63 
64 #include <xen/xen-os.h>
65 #include <xen/hypervisor.h>
66 #include <xen/xen_intr.h>
67 #include <xen/gnttab.h>
68 #include <contrib/xen/memory.h>
69 #include <contrib/xen/io/netif.h>
70 #include <xen/xenbus/xenbusvar.h>
71 
72 #include <machine/bus.h>
73 
74 #include "xenbus_if.h"
75 
76 /* Features supported by all backends.  TSO and LRO can be negotiated */
77 #define XN_CSUM_FEATURES	(CSUM_TCP | CSUM_UDP)
78 
79 #define NET_TX_RING_SIZE __CONST_RING_SIZE(netif_tx, PAGE_SIZE)
80 #define NET_RX_RING_SIZE __CONST_RING_SIZE(netif_rx, PAGE_SIZE)
81 
82 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
83 
84 /*
85  * Should the driver do LRO on the RX end
86  *  this can be toggled on the fly, but the
87  *  interface must be reset (down/up) for it
88  *  to take effect.
89  */
90 static int xn_enable_lro = 1;
91 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
92 
93 /*
94  * Number of pairs of queues.
95  */
96 static unsigned long xn_num_queues = 4;
97 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
98 
99 /**
100  * \brief The maximum allowed data fragments in a single transmit
101  *        request.
102  *
103  * This limit is imposed by the backend driver.  We assume here that
104  * we are dealing with a Linux driver domain and have set our limit
105  * to mirror the Linux MAX_SKB_FRAGS constant.
106  */
107 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
108 
109 #define RX_COPY_THRESHOLD 256
110 
111 #define net_ratelimit() 0
112 
113 struct netfront_rxq;
114 struct netfront_txq;
115 struct netfront_info;
116 struct netfront_rx_info;
117 
118 static void xn_txeof(struct netfront_txq *);
119 static void xn_rxeof(struct netfront_rxq *);
120 static void xn_alloc_rx_buffers(struct netfront_rxq *);
121 static void xn_alloc_rx_buffers_callout(void *arg);
122 
123 static void xn_release_rx_bufs(struct netfront_rxq *);
124 static void xn_release_tx_bufs(struct netfront_txq *);
125 
126 static void xn_rxq_intr(struct netfront_rxq *);
127 static void xn_txq_intr(struct netfront_txq *);
128 static void xn_intr(void *);
129 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
130 static int xn_ioctl(if_t, u_long, caddr_t);
131 static void xn_ifinit_locked(struct netfront_info *);
132 static void xn_ifinit(void *);
133 static void xn_stop(struct netfront_info *);
134 static void xn_query_features(struct netfront_info *np);
135 static int xn_configure_features(struct netfront_info *np);
136 static void netif_free(struct netfront_info *info);
137 static int netfront_detach(device_t dev);
138 
139 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
140 static int xn_txq_mq_start(if_t, struct mbuf *);
141 
142 static int talk_to_backend(device_t dev, struct netfront_info *info);
143 static int create_netdev(device_t dev);
144 static void netif_disconnect_backend(struct netfront_info *info);
145 static int setup_device(device_t dev, struct netfront_info *info,
146     unsigned long);
147 static int xn_ifmedia_upd(if_t ifp);
148 static void xn_ifmedia_sts(if_t ifp, struct ifmediareq *ifmr);
149 
150 static int xn_connect(struct netfront_info *);
151 static void xn_kick_rings(struct netfront_info *);
152 
153 static int xn_get_responses(struct netfront_rxq *,
154     struct netfront_rx_info *, RING_IDX, RING_IDX *,
155     struct mbuf **);
156 
157 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
158 
159 #define INVALID_P2M_ENTRY (~0UL)
160 #define XN_QUEUE_NAME_LEN  8	/* xn{t,r}x_%u, allow for two digits */
161 struct netfront_rxq {
162 	struct netfront_info 	*info;
163 	u_int			id;
164 	char			name[XN_QUEUE_NAME_LEN];
165 	struct mtx		lock;
166 
167 	int			ring_ref;
168 	netif_rx_front_ring_t 	ring;
169 	xen_intr_handle_t	xen_intr_handle;
170 
171 	grant_ref_t 		gref_head;
172 	grant_ref_t 		grant_ref[NET_RX_RING_SIZE + 1];
173 
174 	struct mbuf		*mbufs[NET_RX_RING_SIZE + 1];
175 
176 	struct lro_ctrl		lro;
177 
178 	struct callout		rx_refill;
179 };
180 
181 struct netfront_txq {
182 	struct netfront_info 	*info;
183 	u_int 			id;
184 	char			name[XN_QUEUE_NAME_LEN];
185 	struct mtx		lock;
186 
187 	int			ring_ref;
188 	netif_tx_front_ring_t	ring;
189 	xen_intr_handle_t 	xen_intr_handle;
190 
191 	grant_ref_t		gref_head;
192 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
193 
194 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
195 	int			mbufs_cnt;
196 	struct buf_ring		*br;
197 
198 	struct taskqueue 	*tq;
199 	struct task       	defrtask;
200 
201 	bus_dma_segment_t	segs[MAX_TX_REQ_FRAGS];
202 	struct mbuf_xennet {
203 		struct m_tag 	tag;
204 		bus_dma_tag_t	dma_tag;
205 		bus_dmamap_t	dma_map;
206 		struct netfront_txq *txq;
207 		SLIST_ENTRY(mbuf_xennet) next;
208 		u_int 		count;
209 	}			xennet_tag[NET_TX_RING_SIZE + 1];
210 	SLIST_HEAD(, mbuf_xennet) tags;
211 
212 	bool			full;
213 };
214 
215 struct netfront_info {
216 	if_t			xn_ifp;
217 
218 	struct mtx   		sc_lock;
219 
220 	u_int  num_queues;
221 	struct netfront_rxq 	*rxq;
222 	struct netfront_txq 	*txq;
223 
224 	u_int			carrier;
225 	u_int			maxfrags;
226 
227 	device_t		xbdev;
228 	uint8_t			mac[ETHER_ADDR_LEN];
229 
230 	int			xn_if_flags;
231 
232 	struct ifmedia		sc_media;
233 
234 	bus_dma_tag_t		dma_tag;
235 
236 	bool			xn_reset;
237 };
238 
239 struct netfront_rx_info {
240 	struct netif_rx_response rx;
241 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
242 };
243 
244 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
245 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
246 
247 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
248 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
249 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
250 
251 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
252 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
253 
254 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
255 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
256 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
257 
258 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
259 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
260 #define netfront_carrier_ok(netif)	((netif)->carrier)
261 
262 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
263 
264 static inline void
265 add_id_to_freelist(struct mbuf **list, uintptr_t id)
266 {
267 
268 	KASSERT(id != 0,
269 		("%s: the head item (0) must always be free.", __func__));
270 	list[id] = list[0];
271 	list[0]  = (struct mbuf *)id;
272 }
273 
274 static inline unsigned short
275 get_id_from_freelist(struct mbuf **list)
276 {
277 	uintptr_t id;
278 
279 	id = (uintptr_t)list[0];
280 	KASSERT(id != 0,
281 		("%s: the head item (0) must always remain free.", __func__));
282 	list[0] = list[id];
283 	return (id);
284 }
285 
286 static inline int
287 xn_rxidx(RING_IDX idx)
288 {
289 
290 	return idx & (NET_RX_RING_SIZE - 1);
291 }
292 
293 static inline struct mbuf *
294 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
295 {
296 	int i;
297 	struct mbuf *m;
298 
299 	i = xn_rxidx(ri);
300 	m = rxq->mbufs[i];
301 	rxq->mbufs[i] = NULL;
302 	return (m);
303 }
304 
305 static inline grant_ref_t
306 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
307 {
308 	int i = xn_rxidx(ri);
309 	grant_ref_t ref = rxq->grant_ref[i];
310 
311 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
312 	rxq->grant_ref[i] = GRANT_REF_INVALID;
313 	return (ref);
314 }
315 
316 #define MTAG_COOKIE 1218492000
317 #define MTAG_XENNET 0
318 
319 static void mbuf_grab(struct mbuf *m)
320 {
321 	struct mbuf_xennet *ref;
322 
323 	ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
324 	    MTAG_XENNET, NULL);
325 	KASSERT(ref != NULL, ("Cannot find refcount"));
326 	ref->count++;
327 }
328 
329 static void mbuf_release(struct mbuf *m)
330 {
331 	struct mbuf_xennet *ref;
332 
333 	ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
334 	    MTAG_XENNET, NULL);
335 	KASSERT(ref != NULL, ("Cannot find refcount"));
336 	KASSERT(ref->count > 0, ("Invalid reference count"));
337 
338 	if (--ref->count == 0)
339 		m_freem(m);
340 }
341 
342 static void tag_free(struct m_tag *t)
343 {
344 	struct mbuf_xennet *ref = (struct mbuf_xennet *)t;
345 
346 	KASSERT(ref->count == 0, ("Free mbuf tag with pending refcnt"));
347 	bus_dmamap_sync(ref->dma_tag, ref->dma_map, BUS_DMASYNC_POSTWRITE);
348 	bus_dmamap_destroy(ref->dma_tag, ref->dma_map);
349 	SLIST_INSERT_HEAD(&ref->txq->tags, ref, next);
350 }
351 
352 #define IPRINTK(fmt, args...) \
353     printf("[XEN] " fmt, ##args)
354 #ifdef INVARIANTS
355 #define WPRINTK(fmt, args...) \
356     printf("[XEN] " fmt, ##args)
357 #else
358 #define WPRINTK(fmt, args...)
359 #endif
360 #ifdef DEBUG
361 #define DPRINTK(fmt, args...) \
362     printf("[XEN] %s: " fmt, __func__, ##args)
363 #else
364 #define DPRINTK(fmt, args...)
365 #endif
366 
367 /**
368  * Read the 'mac' node at the given device's node in the store, and parse that
369  * as colon-separated octets, placing result the given mac array.  mac must be
370  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
371  * Return 0 on success, or errno on error.
372  */
373 static int
374 xen_net_read_mac(device_t dev, uint8_t mac[])
375 {
376 	int error, i;
377 	char *s, *e, *macstr;
378 	const char *path;
379 
380 	path = xenbus_get_node(dev);
381 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
382 	if (error == ENOENT) {
383 		/*
384 		 * Deal with missing mac XenStore nodes on devices with
385 		 * HVM emulation (the 'ioemu' configuration attribute)
386 		 * enabled.
387 		 *
388 		 * The HVM emulator may execute in a stub device model
389 		 * domain which lacks the permission, only given to Dom0,
390 		 * to update the guest's XenStore tree.  For this reason,
391 		 * the HVM emulator doesn't even attempt to write the
392 		 * front-side mac node, even when operating in Dom0.
393 		 * However, there should always be a mac listed in the
394 		 * backend tree.  Fallback to this version if our query
395 		 * of the front side XenStore location doesn't find
396 		 * anything.
397 		 */
398 		path = xenbus_get_otherend_path(dev);
399 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
400 	}
401 	if (error != 0) {
402 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
403 		return (error);
404 	}
405 
406 	s = macstr;
407 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
408 		mac[i] = strtoul(s, &e, 16);
409 		if (s == e || (e[0] != ':' && e[0] != 0)) {
410 			free(macstr, M_XENBUS);
411 			return (ENOENT);
412 		}
413 		s = &e[1];
414 	}
415 	free(macstr, M_XENBUS);
416 	return (0);
417 }
418 
419 /**
420  * Entry point to this code when a new device is created.  Allocate the basic
421  * structures and the ring buffers for communication with the backend, and
422  * inform the backend of the appropriate details for those.  Switch to
423  * Connected state.
424  */
425 static int
426 netfront_probe(device_t dev)
427 {
428 
429 	if (xen_pv_nics_disabled())
430 		return (ENXIO);
431 
432 	if (!strcmp(xenbus_get_type(dev), "vif")) {
433 		device_set_desc(dev, "Virtual Network Interface");
434 		return (0);
435 	}
436 
437 	return (ENXIO);
438 }
439 
440 static int
441 netfront_attach(device_t dev)
442 {
443 	int err;
444 
445 	err = create_netdev(dev);
446 	if (err != 0) {
447 		xenbus_dev_fatal(dev, err, "creating netdev");
448 		return (err);
449 	}
450 
451 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
452 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
453 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
454 	    &xn_enable_lro, 0, "Large Receive Offload");
455 
456 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
457 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
458 	    OID_AUTO, "num_queues", CTLFLAG_RD,
459 	    &xn_num_queues, "Number of pairs of queues");
460 
461 	return (0);
462 }
463 
464 static int
465 netfront_suspend(device_t dev)
466 {
467 	struct netfront_info *np = device_get_softc(dev);
468 	u_int i;
469 
470 	for (i = 0; i < np->num_queues; i++) {
471 		XN_RX_LOCK(&np->rxq[i]);
472 		XN_TX_LOCK(&np->txq[i]);
473 	}
474 	netfront_carrier_off(np);
475 	for (i = 0; i < np->num_queues; i++) {
476 		XN_RX_UNLOCK(&np->rxq[i]);
477 		XN_TX_UNLOCK(&np->txq[i]);
478 	}
479 	return (0);
480 }
481 
482 /**
483  * We are reconnecting to the backend, due to a suspend/resume, or a backend
484  * driver restart.  We tear down our netif structure and recreate it, but
485  * leave the device-layer structures intact so that this is transparent to the
486  * rest of the kernel.
487  */
488 static int
489 netfront_resume(device_t dev)
490 {
491 	struct netfront_info *info = device_get_softc(dev);
492 	u_int i;
493 
494 	if (xen_suspend_cancelled) {
495 		for (i = 0; i < info->num_queues; i++) {
496 			XN_RX_LOCK(&info->rxq[i]);
497 			XN_TX_LOCK(&info->txq[i]);
498 		}
499 		netfront_carrier_on(info);
500 		for (i = 0; i < info->num_queues; i++) {
501 			XN_RX_UNLOCK(&info->rxq[i]);
502 			XN_TX_UNLOCK(&info->txq[i]);
503 		}
504 		return (0);
505 	}
506 
507 	netif_disconnect_backend(info);
508 	return (0);
509 }
510 
511 static int
512 write_queue_xenstore_keys(device_t dev,
513     struct netfront_rxq *rxq,
514     struct netfront_txq *txq,
515     struct xs_transaction *xst, bool hierarchy)
516 {
517 	int err;
518 	const char *message;
519 	const char *node = xenbus_get_node(dev);
520 	char *path;
521 	size_t path_size;
522 
523 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
524 	/* Split event channel support is not yet there. */
525 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
526 	    ("Split event channels are not supported"));
527 
528 	if (hierarchy) {
529 		path_size = strlen(node) + 10;
530 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
531 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
532 	} else {
533 		path_size = strlen(node) + 1;
534 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
535 		snprintf(path, path_size, "%s", node);
536 	}
537 
538 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
539 	if (err != 0) {
540 		message = "writing tx ring-ref";
541 		goto error;
542 	}
543 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
544 	if (err != 0) {
545 		message = "writing rx ring-ref";
546 		goto error;
547 	}
548 	err = xs_printf(*xst, path, "event-channel", "%u",
549 	    xen_intr_port(rxq->xen_intr_handle));
550 	if (err != 0) {
551 		message = "writing event-channel";
552 		goto error;
553 	}
554 
555 	free(path, M_DEVBUF);
556 
557 	return (0);
558 
559 error:
560 	free(path, M_DEVBUF);
561 	xenbus_dev_fatal(dev, err, "%s", message);
562 
563 	return (err);
564 }
565 
566 /* Common code used when first setting up, and when resuming. */
567 static int
568 talk_to_backend(device_t dev, struct netfront_info *info)
569 {
570 	const char *message;
571 	struct xs_transaction xst;
572 	const char *node = xenbus_get_node(dev);
573 	int err;
574 	unsigned long num_queues, max_queues = 0;
575 	unsigned int i;
576 
577 	err = xen_net_read_mac(dev, info->mac);
578 	if (err != 0) {
579 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
580 		goto out;
581 	}
582 
583 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
584 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
585 	if (err != 0)
586 		max_queues = 1;
587 	num_queues = xn_num_queues;
588 	if (num_queues > max_queues)
589 		num_queues = max_queues;
590 
591 	err = setup_device(dev, info, num_queues);
592 	if (err != 0) {
593 		xenbus_dev_fatal(dev, err, "setup device");
594 		goto out;
595 	}
596 
597  again:
598 	err = xs_transaction_start(&xst);
599 	if (err != 0) {
600 		xenbus_dev_fatal(dev, err, "starting transaction");
601 		goto free;
602 	}
603 
604 	if (info->num_queues == 1) {
605 		err = write_queue_xenstore_keys(dev, &info->rxq[0],
606 		    &info->txq[0], &xst, false);
607 		if (err != 0)
608 			goto abort_transaction_no_def_error;
609 	} else {
610 		err = xs_printf(xst, node, "multi-queue-num-queues",
611 		    "%u", info->num_queues);
612 		if (err != 0) {
613 			message = "writing multi-queue-num-queues";
614 			goto abort_transaction;
615 		}
616 
617 		for (i = 0; i < info->num_queues; i++) {
618 			err = write_queue_xenstore_keys(dev, &info->rxq[i],
619 			    &info->txq[i], &xst, true);
620 			if (err != 0)
621 				goto abort_transaction_no_def_error;
622 		}
623 	}
624 
625 	err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
626 	if (err != 0) {
627 		message = "writing request-rx-copy";
628 		goto abort_transaction;
629 	}
630 	err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
631 	if (err != 0) {
632 		message = "writing feature-rx-notify";
633 		goto abort_transaction;
634 	}
635 	err = xs_printf(xst, node, "feature-sg", "%d", 1);
636 	if (err != 0) {
637 		message = "writing feature-sg";
638 		goto abort_transaction;
639 	}
640 	if ((if_getcapenable(info->xn_ifp) & IFCAP_LRO) != 0) {
641 		err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
642 		if (err != 0) {
643 			message = "writing feature-gso-tcpv4";
644 			goto abort_transaction;
645 		}
646 	}
647 	if ((if_getcapenable(info->xn_ifp) & IFCAP_RXCSUM) == 0) {
648 		err = xs_printf(xst, node, "feature-no-csum-offload", "%d", 1);
649 		if (err != 0) {
650 			message = "writing feature-no-csum-offload";
651 			goto abort_transaction;
652 		}
653 	}
654 
655 	err = xs_transaction_end(xst, 0);
656 	if (err != 0) {
657 		if (err == EAGAIN)
658 			goto again;
659 		xenbus_dev_fatal(dev, err, "completing transaction");
660 		goto free;
661 	}
662 
663 	return 0;
664 
665  abort_transaction:
666 	xenbus_dev_fatal(dev, err, "%s", message);
667  abort_transaction_no_def_error:
668 	xs_transaction_end(xst, 1);
669  free:
670 	netif_free(info);
671  out:
672 	return (err);
673 }
674 
675 static void
676 xn_rxq_intr(struct netfront_rxq *rxq)
677 {
678 
679 	XN_RX_LOCK(rxq);
680 	xn_rxeof(rxq);
681 	XN_RX_UNLOCK(rxq);
682 }
683 
684 static void
685 xn_txq_start(struct netfront_txq *txq)
686 {
687 	struct netfront_info *np = txq->info;
688 	if_t ifp = np->xn_ifp;
689 
690 	XN_TX_LOCK_ASSERT(txq);
691 	if (!drbr_empty(ifp, txq->br))
692 		xn_txq_mq_start_locked(txq, NULL);
693 }
694 
695 static void
696 xn_txq_intr(struct netfront_txq *txq)
697 {
698 
699 	XN_TX_LOCK(txq);
700 	if (RING_HAS_UNCONSUMED_RESPONSES(&txq->ring))
701 		xn_txeof(txq);
702 	xn_txq_start(txq);
703 	XN_TX_UNLOCK(txq);
704 }
705 
706 static void
707 xn_txq_tq_deferred(void *xtxq, int pending)
708 {
709 	struct netfront_txq *txq = xtxq;
710 
711 	XN_TX_LOCK(txq);
712 	xn_txq_start(txq);
713 	XN_TX_UNLOCK(txq);
714 }
715 
716 static void
717 disconnect_rxq(struct netfront_rxq *rxq)
718 {
719 
720 	xn_release_rx_bufs(rxq);
721 	gnttab_free_grant_references(rxq->gref_head);
722 	if (rxq->ring_ref != GRANT_REF_INVALID) {
723 		gnttab_end_foreign_access(rxq->ring_ref, NULL);
724 		rxq->ring_ref = GRANT_REF_INVALID;
725 	}
726 	/*
727 	 * No split event channel support at the moment, handle will
728 	 * be unbound in tx. So no need to call xen_intr_unbind here,
729 	 * but we do want to reset the handler to 0.
730 	 */
731 	rxq->xen_intr_handle = 0;
732 }
733 
734 static void
735 destroy_rxq(struct netfront_rxq *rxq)
736 {
737 
738 	callout_drain(&rxq->rx_refill);
739 	free(rxq->ring.sring, M_DEVBUF);
740 	rxq->ring.sring = NULL;
741 }
742 
743 static void
744 destroy_rxqs(struct netfront_info *np)
745 {
746 	int i;
747 
748 	for (i = 0; i < np->num_queues; i++)
749 		destroy_rxq(&np->rxq[i]);
750 
751 	free(np->rxq, M_DEVBUF);
752 	np->rxq = NULL;
753 }
754 
755 static int
756 setup_rxqs(device_t dev, struct netfront_info *info,
757 	   unsigned long num_queues)
758 {
759 	int q, i;
760 	int error;
761 	netif_rx_sring_t *rxs;
762 	struct netfront_rxq *rxq;
763 
764 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
765 	    M_DEVBUF, M_WAITOK|M_ZERO);
766 
767 	for (q = 0; q < num_queues; q++) {
768 		rxq = &info->rxq[q];
769 
770 		rxq->id = q;
771 		rxq->info = info;
772 
773 		rxq->gref_head = GNTTAB_LIST_END;
774 		rxq->ring_ref = GRANT_REF_INVALID;
775 		rxq->ring.sring = NULL;
776 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
777 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
778 		    MTX_DEF);
779 
780 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
781 			rxq->mbufs[i] = NULL;
782 			rxq->grant_ref[i] = GRANT_REF_INVALID;
783 		}
784 
785 		/* Start resources allocation */
786 
787 		if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
788 		    &rxq->gref_head) != 0) {
789 			device_printf(dev, "allocating rx gref");
790 			error = ENOMEM;
791 			goto fail;
792 		}
793 
794 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
795 		    M_WAITOK|M_ZERO);
796 		SHARED_RING_INIT(rxs);
797 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
798 
799 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
800 		    &rxq->ring_ref);
801 		if (error != 0) {
802 			device_printf(dev, "granting rx ring page");
803 			goto fail_grant_ring;
804 		}
805 
806 		callout_init(&rxq->rx_refill, 1);
807 	}
808 
809 	return (0);
810 
811 fail_grant_ring:
812 	gnttab_free_grant_references(rxq->gref_head);
813 	free(rxq->ring.sring, M_DEVBUF);
814 fail:
815 	for (; q >= 0; q--) {
816 		disconnect_rxq(&info->rxq[q]);
817 		destroy_rxq(&info->rxq[q]);
818 	}
819 
820 	free(info->rxq, M_DEVBUF);
821 	return (error);
822 }
823 
824 static void
825 disconnect_txq(struct netfront_txq *txq)
826 {
827 
828 	xn_release_tx_bufs(txq);
829 	gnttab_free_grant_references(txq->gref_head);
830 	if (txq->ring_ref != GRANT_REF_INVALID) {
831 		gnttab_end_foreign_access(txq->ring_ref, NULL);
832 		txq->ring_ref = GRANT_REF_INVALID;
833 	}
834 	xen_intr_unbind(&txq->xen_intr_handle);
835 }
836 
837 static void
838 destroy_txq(struct netfront_txq *txq)
839 {
840 	unsigned int i;
841 
842 	free(txq->ring.sring, M_DEVBUF);
843 	txq->ring.sring = NULL;
844 	buf_ring_free(txq->br, M_DEVBUF);
845 	txq->br = NULL;
846 	if (txq->tq) {
847 		taskqueue_drain_all(txq->tq);
848 		taskqueue_free(txq->tq);
849 		txq->tq = NULL;
850 	}
851 
852 	for (i = 0; i <= NET_TX_RING_SIZE; i++) {
853 		bus_dmamap_destroy(txq->info->dma_tag,
854 		    txq->xennet_tag[i].dma_map);
855 		txq->xennet_tag[i].dma_map = NULL;
856 	}
857 }
858 
859 static void
860 destroy_txqs(struct netfront_info *np)
861 {
862 	int i;
863 
864 	for (i = 0; i < np->num_queues; i++)
865 		destroy_txq(&np->txq[i]);
866 
867 	free(np->txq, M_DEVBUF);
868 	np->txq = NULL;
869 }
870 
871 static int
872 setup_txqs(device_t dev, struct netfront_info *info,
873 	   unsigned long num_queues)
874 {
875 	int q, i;
876 	int error;
877 	netif_tx_sring_t *txs;
878 	struct netfront_txq *txq;
879 
880 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
881 	    M_DEVBUF, M_WAITOK|M_ZERO);
882 
883 	for (q = 0; q < num_queues; q++) {
884 		txq = &info->txq[q];
885 
886 		txq->id = q;
887 		txq->info = info;
888 
889 		txq->gref_head = GNTTAB_LIST_END;
890 		txq->ring_ref = GRANT_REF_INVALID;
891 		txq->ring.sring = NULL;
892 
893 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
894 
895 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
896 		    MTX_DEF);
897 		SLIST_INIT(&txq->tags);
898 
899 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
900 			txq->mbufs[i] = (void *) ((u_long) i+1);
901 			txq->grant_ref[i] = GRANT_REF_INVALID;
902 			txq->xennet_tag[i].txq = txq;
903 			txq->xennet_tag[i].dma_tag = info->dma_tag;
904 			error = bus_dmamap_create(info->dma_tag, 0,
905 			    &txq->xennet_tag[i].dma_map);
906 			if (error != 0) {
907 				device_printf(dev,
908 				    "failed to allocate dma map\n");
909 				goto fail;
910 			}
911 			m_tag_setup(&txq->xennet_tag[i].tag,
912 			    MTAG_COOKIE, MTAG_XENNET,
913 			    sizeof(txq->xennet_tag[i]) -
914 			    sizeof(txq->xennet_tag[i].tag));
915 			txq->xennet_tag[i].tag.m_tag_free = &tag_free;
916 			SLIST_INSERT_HEAD(&txq->tags, &txq->xennet_tag[i],
917 			    next);
918 		}
919 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
920 
921 		/* Start resources allocation. */
922 
923 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
924 		    &txq->gref_head) != 0) {
925 			device_printf(dev, "failed to allocate tx grant refs\n");
926 			error = ENOMEM;
927 			goto fail;
928 		}
929 
930 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
931 		    M_WAITOK|M_ZERO);
932 		SHARED_RING_INIT(txs);
933 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
934 
935 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
936 		    &txq->ring_ref);
937 		if (error != 0) {
938 			device_printf(dev, "failed to grant tx ring\n");
939 			goto fail_grant_ring;
940 		}
941 
942 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
943 		    M_WAITOK, &txq->lock);
944 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
945 
946 		txq->tq = taskqueue_create(txq->name, M_WAITOK,
947 		    taskqueue_thread_enqueue, &txq->tq);
948 
949 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
950 		    "%s txq %d", device_get_nameunit(dev), txq->id);
951 		if (error != 0) {
952 			device_printf(dev, "failed to start tx taskq %d\n",
953 			    txq->id);
954 			goto fail_start_thread;
955 		}
956 
957 		error = xen_intr_alloc_and_bind_local_port(dev,
958 		    xenbus_get_otherend_id(dev), /* filter */ NULL, xn_intr,
959 		    &info->txq[q], INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
960 		    &txq->xen_intr_handle);
961 
962 		if (error != 0) {
963 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
964 			goto fail_bind_port;
965 		}
966 	}
967 
968 	return (0);
969 
970 fail_bind_port:
971 	taskqueue_drain_all(txq->tq);
972 fail_start_thread:
973 	buf_ring_free(txq->br, M_DEVBUF);
974 	taskqueue_free(txq->tq);
975 	gnttab_end_foreign_access(txq->ring_ref, NULL);
976 fail_grant_ring:
977 	gnttab_free_grant_references(txq->gref_head);
978 	free(txq->ring.sring, M_DEVBUF);
979 fail:
980 	for (; q >= 0; q--) {
981 		disconnect_txq(&info->txq[q]);
982 		destroy_txq(&info->txq[q]);
983 	}
984 
985 	free(info->txq, M_DEVBUF);
986 	return (error);
987 }
988 
989 static int
990 setup_device(device_t dev, struct netfront_info *info,
991     unsigned long num_queues)
992 {
993 	int error;
994 	int q;
995 
996 	if (info->txq)
997 		destroy_txqs(info);
998 
999 	if (info->rxq)
1000 		destroy_rxqs(info);
1001 
1002 	info->num_queues = 0;
1003 
1004 	error = setup_rxqs(dev, info, num_queues);
1005 	if (error != 0)
1006 		goto out;
1007 	error = setup_txqs(dev, info, num_queues);
1008 	if (error != 0)
1009 		goto out;
1010 
1011 	info->num_queues = num_queues;
1012 
1013 	/* No split event channel at the moment. */
1014 	for (q = 0; q < num_queues; q++)
1015 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
1016 
1017 	return (0);
1018 
1019 out:
1020 	KASSERT(error != 0, ("Error path taken without providing an error code"));
1021 	return (error);
1022 }
1023 
1024 #ifdef INET
1025 static u_int
1026 netfront_addr_cb(void *arg, struct ifaddr *a, u_int count)
1027 {
1028 	arp_ifinit((if_t)arg, a);
1029 	return (1);
1030 }
1031 /**
1032  * If this interface has an ipv4 address, send an arp for it. This
1033  * helps to get the network going again after migrating hosts.
1034  */
1035 static void
1036 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
1037 {
1038 	if_t ifp;
1039 
1040 	ifp = info->xn_ifp;
1041 	if_foreach_addr_type(ifp, AF_INET, netfront_addr_cb, ifp);
1042 }
1043 #endif
1044 
1045 /**
1046  * Callback received when the backend's state changes.
1047  */
1048 static void
1049 netfront_backend_changed(device_t dev, XenbusState newstate)
1050 {
1051 	struct netfront_info *sc = device_get_softc(dev);
1052 
1053 	DPRINTK("newstate=%d\n", newstate);
1054 
1055 	CURVNET_SET(if_getvnet(sc->xn_ifp));
1056 
1057 	switch (newstate) {
1058 	case XenbusStateInitialising:
1059 	case XenbusStateInitialised:
1060 	case XenbusStateUnknown:
1061 	case XenbusStateReconfigured:
1062 	case XenbusStateReconfiguring:
1063 		break;
1064 	case XenbusStateInitWait:
1065 		if (xenbus_get_state(dev) != XenbusStateInitialising)
1066 			break;
1067 		if (xn_connect(sc) != 0)
1068 			break;
1069 		/* Switch to connected state before kicking the rings. */
1070 		xenbus_set_state(sc->xbdev, XenbusStateConnected);
1071 		xn_kick_rings(sc);
1072 		break;
1073 	case XenbusStateClosing:
1074 		xenbus_set_state(dev, XenbusStateClosed);
1075 		break;
1076 	case XenbusStateClosed:
1077 		if (sc->xn_reset) {
1078 			netif_disconnect_backend(sc);
1079 			xenbus_set_state(dev, XenbusStateInitialising);
1080 			sc->xn_reset = false;
1081 		}
1082 		break;
1083 	case XenbusStateConnected:
1084 #ifdef INET
1085 		netfront_send_fake_arp(dev, sc);
1086 #endif
1087 		break;
1088 	}
1089 
1090 	CURVNET_RESTORE();
1091 }
1092 
1093 /**
1094  * \brief Verify that there is sufficient space in the Tx ring
1095  *        buffer for a maximally sized request to be enqueued.
1096  *
1097  * A transmit request requires a transmit descriptor for each packet
1098  * fragment, plus up to 2 entries for "options" (e.g. TSO).
1099  */
1100 static inline int
1101 xn_tx_slot_available(struct netfront_txq *txq)
1102 {
1103 
1104 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1105 }
1106 
1107 static void
1108 xn_release_tx_bufs(struct netfront_txq *txq)
1109 {
1110 	int i;
1111 
1112 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1113 		struct mbuf *m;
1114 
1115 		m = txq->mbufs[i];
1116 
1117 		/*
1118 		 * We assume that no kernel addresses are
1119 		 * less than NET_TX_RING_SIZE.  Any entry
1120 		 * in the table that is below this number
1121 		 * must be an index from free-list tracking.
1122 		 */
1123 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1124 			continue;
1125 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1126 		gnttab_release_grant_reference(&txq->gref_head,
1127 		    txq->grant_ref[i]);
1128 		txq->grant_ref[i] = GRANT_REF_INVALID;
1129 		add_id_to_freelist(txq->mbufs, i);
1130 		txq->mbufs_cnt--;
1131 		if (txq->mbufs_cnt < 0) {
1132 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1133 		}
1134 		mbuf_release(m);
1135 	}
1136 }
1137 
1138 static struct mbuf *
1139 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1140 {
1141 	struct mbuf *m;
1142 
1143 	m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1144 	if (m == NULL)
1145 		return NULL;
1146 	m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1147 
1148 	return (m);
1149 }
1150 
1151 static void
1152 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1153 {
1154 	RING_IDX req_prod;
1155 	int notify;
1156 
1157 	XN_RX_LOCK_ASSERT(rxq);
1158 
1159 	if (__predict_false(rxq->info->carrier == 0))
1160 		return;
1161 
1162 	for (req_prod = rxq->ring.req_prod_pvt;
1163 	     req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1164 	     req_prod++) {
1165 		struct mbuf *m;
1166 		unsigned short id;
1167 		grant_ref_t ref;
1168 		struct netif_rx_request *req;
1169 		unsigned long pfn;
1170 
1171 		m = xn_alloc_one_rx_buffer(rxq);
1172 		if (m == NULL)
1173 			break;
1174 
1175 		id = xn_rxidx(req_prod);
1176 
1177 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1178 		rxq->mbufs[id] = m;
1179 
1180 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1181 		KASSERT(ref != GNTTAB_LIST_END,
1182 		    ("reserved grant references exhuasted"));
1183 		rxq->grant_ref[id] = ref;
1184 
1185 		pfn = atop(vtophys(mtod(m, vm_offset_t)));
1186 		req = RING_GET_REQUEST(&rxq->ring, req_prod);
1187 
1188 		gnttab_grant_foreign_access_ref(ref,
1189 		    xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1190 		req->id = id;
1191 		req->gref = ref;
1192 	}
1193 
1194 	rxq->ring.req_prod_pvt = req_prod;
1195 
1196 	/* Not enough requests? Try again later. */
1197 	if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1198 		callout_reset_curcpu(&rxq->rx_refill, hz/10,
1199 		    xn_alloc_rx_buffers_callout, rxq);
1200 		return;
1201 	}
1202 
1203 	wmb();		/* barrier so backend seens requests */
1204 
1205 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1206 	if (notify)
1207 		xen_intr_signal(rxq->xen_intr_handle);
1208 }
1209 
1210 static void xn_alloc_rx_buffers_callout(void *arg)
1211 {
1212 	struct netfront_rxq *rxq;
1213 
1214 	rxq = (struct netfront_rxq *)arg;
1215 	XN_RX_LOCK(rxq);
1216 	xn_alloc_rx_buffers(rxq);
1217 	XN_RX_UNLOCK(rxq);
1218 }
1219 
1220 static void
1221 xn_release_rx_bufs(struct netfront_rxq *rxq)
1222 {
1223 	int i,  ref;
1224 	struct mbuf *m;
1225 
1226 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1227 		m = rxq->mbufs[i];
1228 
1229 		if (m == NULL)
1230 			continue;
1231 
1232 		ref = rxq->grant_ref[i];
1233 		if (ref == GRANT_REF_INVALID)
1234 			continue;
1235 
1236 		gnttab_end_foreign_access_ref(ref);
1237 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1238 		rxq->mbufs[i] = NULL;
1239 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1240 		m_freem(m);
1241 	}
1242 }
1243 
1244 static void
1245 xn_rxeof(struct netfront_rxq *rxq)
1246 {
1247 	if_t ifp;
1248 	struct netfront_info *np = rxq->info;
1249 #if (defined(INET) || defined(INET6))
1250 	struct lro_ctrl *lro = &rxq->lro;
1251 #endif
1252 	struct netfront_rx_info rinfo;
1253 	struct netif_rx_response *rx = &rinfo.rx;
1254 	struct netif_extra_info *extras = rinfo.extras;
1255 	RING_IDX i, rp;
1256 	struct mbuf *m;
1257 	struct mbufq mbufq_rxq, mbufq_errq;
1258 	int err, work_to_do;
1259 
1260 	XN_RX_LOCK_ASSERT(rxq);
1261 
1262 	if (!netfront_carrier_ok(np))
1263 		return;
1264 
1265 	/* XXX: there should be some sane limit. */
1266 	mbufq_init(&mbufq_errq, INT_MAX);
1267 	mbufq_init(&mbufq_rxq, INT_MAX);
1268 
1269 	ifp = np->xn_ifp;
1270 
1271 	do {
1272 		rp = rxq->ring.sring->rsp_prod;
1273 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1274 
1275 		i = rxq->ring.rsp_cons;
1276 		while ((i != rp)) {
1277 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1278 			memset(extras, 0, sizeof(rinfo.extras));
1279 
1280 			m = NULL;
1281 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1282 
1283 			if (__predict_false(err)) {
1284 				if (m)
1285 					(void )mbufq_enqueue(&mbufq_errq, m);
1286 				if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
1287 				continue;
1288 			}
1289 
1290 			m->m_pkthdr.rcvif = ifp;
1291 			if (rx->flags & NETRXF_data_validated) {
1292 				/*
1293 				 * According to mbuf(9) the correct way to tell
1294 				 * the stack that the checksum of an inbound
1295 				 * packet is correct, without it actually being
1296 				 * present (because the underlying interface
1297 				 * doesn't provide it), is to set the
1298 				 * CSUM_DATA_VALID and CSUM_PSEUDO_HDR flags,
1299 				 * and the csum_data field to 0xffff.
1300 				 */
1301 				m->m_pkthdr.csum_flags |= (CSUM_DATA_VALID
1302 				    | CSUM_PSEUDO_HDR);
1303 				m->m_pkthdr.csum_data = 0xffff;
1304 			}
1305 			if ((rx->flags & NETRXF_extra_info) != 0 &&
1306 			    (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1307 			    XEN_NETIF_EXTRA_TYPE_GSO)) {
1308 				m->m_pkthdr.tso_segsz =
1309 				extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1310 				m->m_pkthdr.csum_flags |= CSUM_TSO;
1311 			}
1312 
1313 			(void )mbufq_enqueue(&mbufq_rxq, m);
1314 		}
1315 
1316 		rxq->ring.rsp_cons = i;
1317 
1318 		xn_alloc_rx_buffers(rxq);
1319 
1320 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1321 	} while (work_to_do);
1322 
1323 	mbufq_drain(&mbufq_errq);
1324 	/*
1325 	 * Process all the mbufs after the remapping is complete.
1326 	 * Break the mbuf chain first though.
1327 	 */
1328 	while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1329 		if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1330 #if (defined(INET) || defined(INET6))
1331 		/* Use LRO if possible */
1332 		if ((if_getcapenable(ifp) & IFCAP_LRO) == 0 ||
1333 		    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1334 			/*
1335 			 * If LRO fails, pass up to the stack
1336 			 * directly.
1337 			 */
1338 			if_input(ifp, m);
1339 		}
1340 #else
1341 		if_input(ifp, m);
1342 #endif
1343 	}
1344 
1345 #if (defined(INET) || defined(INET6))
1346 	/*
1347 	 * Flush any outstanding LRO work
1348 	 */
1349 	tcp_lro_flush_all(lro);
1350 #endif
1351 }
1352 
1353 static void
1354 xn_txeof(struct netfront_txq *txq)
1355 {
1356 	RING_IDX i, prod;
1357 	unsigned short id;
1358 	if_t ifp;
1359 	netif_tx_response_t *txr;
1360 	struct mbuf *m;
1361 	struct netfront_info *np = txq->info;
1362 
1363 	XN_TX_LOCK_ASSERT(txq);
1364 
1365 	if (!netfront_carrier_ok(np))
1366 		return;
1367 
1368 	ifp = np->xn_ifp;
1369 
1370 	do {
1371 		prod = txq->ring.sring->rsp_prod;
1372 		rmb(); /* Ensure we see responses up to 'rp'. */
1373 
1374 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1375 			txr = RING_GET_RESPONSE(&txq->ring, i);
1376 			if (txr->status == NETIF_RSP_NULL)
1377 				continue;
1378 
1379 			if (txr->status != NETIF_RSP_OKAY) {
1380 				printf("%s: WARNING: response is %d!\n",
1381 				       __func__, txr->status);
1382 			}
1383 			id = txr->id;
1384 			m = txq->mbufs[id];
1385 			KASSERT(m != NULL, ("mbuf not found in chain"));
1386 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1387 				("mbuf already on the free list, but we're "
1388 				"trying to free it again!"));
1389 			M_ASSERTVALID(m);
1390 
1391 			if (__predict_false(gnttab_query_foreign_access(
1392 			    txq->grant_ref[id]) != 0)) {
1393 				panic("%s: grant id %u still in use by the "
1394 				    "backend", __func__, id);
1395 			}
1396 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1397 			gnttab_release_grant_reference(
1398 				&txq->gref_head, txq->grant_ref[id]);
1399 			txq->grant_ref[id] = GRANT_REF_INVALID;
1400 
1401 			txq->mbufs[id] = NULL;
1402 			add_id_to_freelist(txq->mbufs, id);
1403 			txq->mbufs_cnt--;
1404 			mbuf_release(m);
1405 			/* Only mark the txq active if we've freed up at least one slot to try */
1406 			if_setdrvflagbits(ifp, 0, IFF_DRV_OACTIVE);
1407 		}
1408 		txq->ring.rsp_cons = prod;
1409 
1410 		/*
1411 		 * Set a new event, then check for race with update of
1412 		 * tx_cons. Note that it is essential to schedule a
1413 		 * callback, no matter how few buffers are pending. Even if
1414 		 * there is space in the transmit ring, higher layers may
1415 		 * be blocked because too much data is outstanding: in such
1416 		 * cases notification from Xen is likely to be the only kick
1417 		 * that we'll get.
1418 		 */
1419 		txq->ring.sring->rsp_event =
1420 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1421 
1422 		mb();
1423 	} while (prod != txq->ring.sring->rsp_prod);
1424 
1425 	if (txq->full &&
1426 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1427 		txq->full = false;
1428 		xn_txq_start(txq);
1429 	}
1430 }
1431 
1432 static void
1433 xn_intr(void *xsc)
1434 {
1435 	struct netfront_txq *txq = xsc;
1436 	struct netfront_info *np = txq->info;
1437 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1438 
1439 	/* kick both tx and rx */
1440 	xn_rxq_intr(rxq);
1441 	xn_txq_intr(txq);
1442 }
1443 
1444 static void
1445 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1446     grant_ref_t ref)
1447 {
1448 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1449 
1450 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1451 	rxq->mbufs[new] = m;
1452 	rxq->grant_ref[new] = ref;
1453 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1454 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1455 	rxq->ring.req_prod_pvt++;
1456 }
1457 
1458 static int
1459 xn_get_extras(struct netfront_rxq *rxq,
1460     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1461 {
1462 	struct netif_extra_info *extra;
1463 
1464 	int err = 0;
1465 
1466 	do {
1467 		struct mbuf *m;
1468 		grant_ref_t ref;
1469 
1470 		if (__predict_false(*cons + 1 == rp)) {
1471 			err = EINVAL;
1472 			break;
1473 		}
1474 
1475 		extra = (struct netif_extra_info *)
1476 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1477 
1478 		if (__predict_false(!extra->type ||
1479 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1480 			err = EINVAL;
1481 		} else {
1482 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1483 		}
1484 
1485 		m = xn_get_rx_mbuf(rxq, *cons);
1486 		ref = xn_get_rx_ref(rxq,  *cons);
1487 		xn_move_rx_slot(rxq, m, ref);
1488 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1489 
1490 	return err;
1491 }
1492 
1493 static int
1494 xn_get_responses(struct netfront_rxq *rxq,
1495     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1496     struct mbuf  **list)
1497 {
1498 	struct netif_rx_response *rx = &rinfo->rx;
1499 	struct netif_extra_info *extras = rinfo->extras;
1500 	struct mbuf *m, *m0, *m_prev;
1501 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1502 	int frags = 1;
1503 	int err = 0;
1504 	u_long ret __diagused;
1505 
1506 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1507 
1508 	if (rx->flags & NETRXF_extra_info) {
1509 		err = xn_get_extras(rxq, extras, rp, cons);
1510 	}
1511 
1512 	if (m0 != NULL) {
1513 		m0->m_pkthdr.len = 0;
1514 		m0->m_next = NULL;
1515 	}
1516 
1517 	for (;;) {
1518 #if 0
1519 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1520 			rx->status, rx->offset, frags);
1521 #endif
1522 		if (__predict_false(rx->status < 0 ||
1523 			rx->offset + rx->status > PAGE_SIZE)) {
1524 			xn_move_rx_slot(rxq, m, ref);
1525 			if (m0 == m)
1526 				m0 = NULL;
1527 			m = NULL;
1528 			err = EINVAL;
1529 			goto next_skip_queue;
1530 		}
1531 
1532 		/*
1533 		 * This definitely indicates a bug, either in this driver or in
1534 		 * the backend driver. In future this should flag the bad
1535 		 * situation to the system controller to reboot the backed.
1536 		 */
1537 		if (ref == GRANT_REF_INVALID) {
1538 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1539 			err = EINVAL;
1540 			goto next;
1541 		}
1542 
1543 		ret = gnttab_end_foreign_access_ref(ref);
1544 		KASSERT(ret, ("Unable to end access to grant references"));
1545 
1546 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1547 
1548 next:
1549 		if (m == NULL)
1550 			break;
1551 
1552 		m->m_len = rx->status;
1553 		m->m_data += rx->offset;
1554 		m0->m_pkthdr.len += rx->status;
1555 
1556 next_skip_queue:
1557 		if (!(rx->flags & NETRXF_more_data))
1558 			break;
1559 
1560 		if (*cons + frags == rp) {
1561 			if (net_ratelimit())
1562 				WPRINTK("Need more frags\n");
1563 			err = ENOENT;
1564 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1565 			       __func__, *cons, frags, rp);
1566 			break;
1567 		}
1568 		/*
1569 		 * Note that m can be NULL, if rx->status < 0 or if
1570 		 * rx->offset + rx->status > PAGE_SIZE above.
1571 		 */
1572 		m_prev = m;
1573 
1574 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1575 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1576 
1577 		/*
1578 		 * m_prev == NULL can happen if rx->status < 0 or if
1579 		 * rx->offset + * rx->status > PAGE_SIZE above.
1580 		 */
1581 		if (m_prev != NULL)
1582 			m_prev->m_next = m;
1583 
1584 		/*
1585 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1586 		 * rx->status > PAGE_SIZE above.
1587 		 */
1588 		if (m0 == NULL)
1589 			m0 = m;
1590 		m->m_next = NULL;
1591 		ref = xn_get_rx_ref(rxq, *cons + frags);
1592 		frags++;
1593 	}
1594 	*list = m0;
1595 	*cons += frags;
1596 
1597 	return (err);
1598 }
1599 
1600 /**
1601  * Given an mbuf chain, make sure we have enough room and then push
1602  * it onto the transmit ring.
1603  */
1604 static int
1605 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1606 {
1607 	struct netfront_info *np = txq->info;
1608 	if_t ifp = np->xn_ifp;
1609 	int otherend_id, error, nfrags;
1610 	bus_dma_segment_t *segs = txq->segs;
1611 	struct mbuf_xennet *tag;
1612 	bus_dmamap_t map;
1613 	unsigned int i;
1614 
1615 	KASSERT(!SLIST_EMPTY(&txq->tags), ("no tags available"));
1616 	tag = SLIST_FIRST(&txq->tags);
1617 	SLIST_REMOVE_HEAD(&txq->tags, next);
1618 	KASSERT(tag->count == 0, ("tag already in-use"));
1619 	map = tag->dma_map;
1620 	error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1621 	    &nfrags, 0);
1622 	if (error == EFBIG || nfrags > np->maxfrags) {
1623 		struct mbuf *m;
1624 
1625 		bus_dmamap_unload(np->dma_tag, map);
1626 		m = m_defrag(m_head, M_NOWAIT);
1627 		if (!m) {
1628 			/*
1629 			 * Defrag failed, so free the mbuf and
1630 			 * therefore drop the packet.
1631 			 */
1632 			SLIST_INSERT_HEAD(&txq->tags, tag, next);
1633 			m_freem(m_head);
1634 			return (EMSGSIZE);
1635 		}
1636 		m_head = m;
1637 		error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1638 		    &nfrags, 0);
1639 		if (error != 0 || nfrags > np->maxfrags) {
1640 			bus_dmamap_unload(np->dma_tag, map);
1641 			SLIST_INSERT_HEAD(&txq->tags, tag, next);
1642 			m_freem(m_head);
1643 			return (error ?: EFBIG);
1644 		}
1645 	} else if (error != 0) {
1646 		SLIST_INSERT_HEAD(&txq->tags, tag, next);
1647 		m_freem(m_head);
1648 		return (error);
1649 	}
1650 
1651 	/**
1652 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1653 	 * of mbufs longer than Linux can handle.  Make sure we don't
1654 	 * pass a too-long chain over to the other side by dropping the
1655 	 * packet.  It doesn't look like there is currently a way to
1656 	 * tell the TCP stack to generate a shorter chain of packets.
1657 	 */
1658 	if (nfrags > MAX_TX_REQ_FRAGS) {
1659 #ifdef DEBUG
1660 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1661 		       "won't be able to handle it, dropping\n",
1662 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1663 #endif
1664 		SLIST_INSERT_HEAD(&txq->tags, tag, next);
1665 		bus_dmamap_unload(np->dma_tag, map);
1666 		m_freem(m_head);
1667 		return (EMSGSIZE);
1668 	}
1669 
1670 	/*
1671 	 * This check should be redundant.  We've already verified that we
1672 	 * have enough slots in the ring to handle a packet of maximum
1673 	 * size, and that our packet is less than the maximum size.  Keep
1674 	 * it in here as an assert for now just to make certain that
1675 	 * chain_cnt is accurate.
1676 	 */
1677 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1678 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1679 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1680                     (int) nfrags, (int) NET_TX_RING_SIZE));
1681 
1682 	/*
1683 	 * Start packing the mbufs in this chain into
1684 	 * the fragment pointers. Stop when we run out
1685 	 * of fragments or hit the end of the mbuf chain.
1686 	 */
1687 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1688 	m_tag_prepend(m_head, &tag->tag);
1689 	for (i = 0; i < nfrags; i++) {
1690 		netif_tx_request_t *tx;
1691 		uintptr_t id;
1692 		grant_ref_t ref;
1693 		u_long mfn; /* XXX Wrong type? */
1694 
1695 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1696 		id = get_id_from_freelist(txq->mbufs);
1697 		if (id == 0)
1698 			panic("%s: was allocated the freelist head!\n",
1699 			    __func__);
1700 		txq->mbufs_cnt++;
1701 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1702 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1703 			    __func__);
1704 		mbuf_grab(m_head);
1705 		txq->mbufs[id] = m_head;
1706 		tx->id = id;
1707 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1708 		KASSERT((short)ref >= 0, ("Negative ref"));
1709 		mfn = atop(segs[i].ds_addr);
1710 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1711 		    mfn, GNTMAP_readonly);
1712 		tx->gref = txq->grant_ref[id] = ref;
1713 		tx->offset = segs[i].ds_addr & PAGE_MASK;
1714 		KASSERT(tx->offset + segs[i].ds_len <= PAGE_SIZE,
1715 		    ("mbuf segment crosses a page boundary"));
1716 		tx->flags = 0;
1717 		if (i == 0) {
1718 			/*
1719 			 * The first fragment has the entire packet
1720 			 * size, subsequent fragments have just the
1721 			 * fragment size. The backend works out the
1722 			 * true size of the first fragment by
1723 			 * subtracting the sizes of the other
1724 			 * fragments.
1725 			 */
1726 			tx->size = m_head->m_pkthdr.len;
1727 
1728 			/*
1729 			 * The first fragment contains the checksum flags
1730 			 * and is optionally followed by extra data for
1731 			 * TSO etc.
1732 			 */
1733 			/**
1734 			 * CSUM_TSO requires checksum offloading.
1735 			 * Some versions of FreeBSD fail to
1736 			 * set CSUM_TCP in the CSUM_TSO case,
1737 			 * so we have to test for CSUM_TSO
1738 			 * explicitly.
1739 			 */
1740 			if (m_head->m_pkthdr.csum_flags
1741 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1742 				tx->flags |= (NETTXF_csum_blank
1743 				    | NETTXF_data_validated);
1744 			}
1745 			if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
1746 				struct netif_extra_info *gso =
1747 					(struct netif_extra_info *)
1748 					RING_GET_REQUEST(&txq->ring,
1749 							 ++txq->ring.req_prod_pvt);
1750 
1751 				tx->flags |= NETTXF_extra_info;
1752 
1753 				gso->u.gso.size = m_head->m_pkthdr.tso_segsz;
1754 				gso->u.gso.type =
1755 					XEN_NETIF_GSO_TYPE_TCPV4;
1756 				gso->u.gso.pad = 0;
1757 				gso->u.gso.features = 0;
1758 
1759 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1760 				gso->flags = 0;
1761 			}
1762 		} else {
1763 			tx->size = segs[i].ds_len;
1764 		}
1765 		if (i != nfrags - 1)
1766 			tx->flags |= NETTXF_more_data;
1767 
1768 		txq->ring.req_prod_pvt++;
1769 	}
1770 	bus_dmamap_sync(np->dma_tag, map, BUS_DMASYNC_PREWRITE);
1771 	BPF_MTAP(ifp, m_head);
1772 
1773 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1774 	if_inc_counter(ifp, IFCOUNTER_OBYTES, m_head->m_pkthdr.len);
1775 	if (m_head->m_flags & M_MCAST)
1776 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
1777 
1778 	xn_txeof(txq);
1779 
1780 	return (0);
1781 }
1782 
1783 /* equivalent of network_open() in Linux */
1784 static void
1785 xn_ifinit_locked(struct netfront_info *np)
1786 {
1787 	if_t ifp;
1788 	int i;
1789 	struct netfront_rxq *rxq;
1790 
1791 	XN_LOCK_ASSERT(np);
1792 
1793 	ifp = np->xn_ifp;
1794 
1795 	if (if_getdrvflags(ifp) & IFF_DRV_RUNNING || !netfront_carrier_ok(np))
1796 		return;
1797 
1798 	xn_stop(np);
1799 
1800 	for (i = 0; i < np->num_queues; i++) {
1801 		rxq = &np->rxq[i];
1802 		XN_RX_LOCK(rxq);
1803 		xn_alloc_rx_buffers(rxq);
1804 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1805 		if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1806 			xn_rxeof(rxq);
1807 		XN_RX_UNLOCK(rxq);
1808 	}
1809 
1810 	if_setdrvflagbits(ifp, IFF_DRV_RUNNING, 0);
1811 	if_setdrvflagbits(ifp, 0, IFF_DRV_OACTIVE);
1812 	if_link_state_change(ifp, LINK_STATE_UP);
1813 }
1814 
1815 static void
1816 xn_ifinit(void *xsc)
1817 {
1818 	struct netfront_info *sc = xsc;
1819 
1820 	XN_LOCK(sc);
1821 	xn_ifinit_locked(sc);
1822 	XN_UNLOCK(sc);
1823 }
1824 
1825 static int
1826 xn_ioctl(if_t ifp, u_long cmd, caddr_t data)
1827 {
1828 	struct netfront_info *sc = if_getsoftc(ifp);
1829 	struct ifreq *ifr = (struct ifreq *) data;
1830 	device_t dev;
1831 #ifdef INET
1832 	struct ifaddr *ifa = (struct ifaddr *)data;
1833 #endif
1834 	int mask, error = 0, reinit;
1835 
1836 	dev = sc->xbdev;
1837 
1838 	switch(cmd) {
1839 	case SIOCSIFADDR:
1840 #ifdef INET
1841 		XN_LOCK(sc);
1842 		if (ifa->ifa_addr->sa_family == AF_INET) {
1843 			if_setflagbits(ifp, IFF_UP, 0);
1844 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
1845 				xn_ifinit_locked(sc);
1846 			arp_ifinit(ifp, ifa);
1847 			XN_UNLOCK(sc);
1848 		} else {
1849 			XN_UNLOCK(sc);
1850 #endif
1851 			error = ether_ioctl(ifp, cmd, data);
1852 #ifdef INET
1853 		}
1854 #endif
1855 		break;
1856 	case SIOCSIFMTU:
1857 		if (if_getmtu(ifp) == ifr->ifr_mtu)
1858 			break;
1859 
1860 		if_setmtu(ifp, ifr->ifr_mtu);
1861 		if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING);
1862 		xn_ifinit(sc);
1863 		break;
1864 	case SIOCSIFFLAGS:
1865 		XN_LOCK(sc);
1866 		if (if_getflags(ifp) & IFF_UP) {
1867 			/*
1868 			 * If only the state of the PROMISC flag changed,
1869 			 * then just use the 'set promisc mode' command
1870 			 * instead of reinitializing the entire NIC. Doing
1871 			 * a full re-init means reloading the firmware and
1872 			 * waiting for it to start up, which may take a
1873 			 * second or two.
1874 			 */
1875 			xn_ifinit_locked(sc);
1876 		} else {
1877 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
1878 				xn_stop(sc);
1879 			}
1880 		}
1881 		sc->xn_if_flags = if_getflags(ifp);
1882 		XN_UNLOCK(sc);
1883 		break;
1884 	case SIOCSIFCAP:
1885 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
1886 		reinit = 0;
1887 
1888 		if (mask & IFCAP_TXCSUM) {
1889 			if_togglecapenable(ifp, IFCAP_TXCSUM);
1890 			if_togglehwassist(ifp, XN_CSUM_FEATURES);
1891 		}
1892 		if (mask & IFCAP_TSO4) {
1893 			if_togglecapenable(ifp, IFCAP_TSO4);
1894 			if_togglehwassist(ifp, CSUM_TSO);
1895 		}
1896 
1897 		if (mask & (IFCAP_RXCSUM | IFCAP_LRO)) {
1898 			/* These Rx features require us to renegotiate. */
1899 			reinit = 1;
1900 
1901 			if (mask & IFCAP_RXCSUM)
1902 				if_togglecapenable(ifp, IFCAP_RXCSUM);
1903 			if (mask & IFCAP_LRO)
1904 				if_togglecapenable(ifp, IFCAP_LRO);
1905 		}
1906 
1907 		if (reinit == 0)
1908 			break;
1909 
1910 		/*
1911 		 * We must reset the interface so the backend picks up the
1912 		 * new features.
1913 		 */
1914 		device_printf(sc->xbdev,
1915 		    "performing interface reset due to feature change\n");
1916 		XN_LOCK(sc);
1917 		netfront_carrier_off(sc);
1918 		sc->xn_reset = true;
1919 		/*
1920 		 * NB: the pending packet queue is not flushed, since
1921 		 * the interface should still support the old options.
1922 		 */
1923 		XN_UNLOCK(sc);
1924 		/*
1925 		 * Delete the xenstore nodes that export features.
1926 		 *
1927 		 * NB: There's a xenbus state called
1928 		 * "XenbusStateReconfiguring", which is what we should set
1929 		 * here. Sadly none of the backends know how to handle it,
1930 		 * and simply disconnect from the frontend, so we will just
1931 		 * switch back to XenbusStateInitialising in order to force
1932 		 * a reconnection.
1933 		 */
1934 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-gso-tcpv4");
1935 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-no-csum-offload");
1936 		xenbus_set_state(dev, XenbusStateClosing);
1937 
1938 		/*
1939 		 * Wait for the frontend to reconnect before returning
1940 		 * from the ioctl. 30s should be more than enough for any
1941 		 * sane backend to reconnect.
1942 		 */
1943 		error = tsleep(sc, 0, "xn_rst", 30*hz);
1944 		break;
1945 	case SIOCADDMULTI:
1946 	case SIOCDELMULTI:
1947 		break;
1948 	case SIOCSIFMEDIA:
1949 	case SIOCGIFMEDIA:
1950 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1951 		break;
1952 	default:
1953 		error = ether_ioctl(ifp, cmd, data);
1954 	}
1955 
1956 	return (error);
1957 }
1958 
1959 static void
1960 xn_stop(struct netfront_info *sc)
1961 {
1962 	if_t ifp;
1963 
1964 	XN_LOCK_ASSERT(sc);
1965 
1966 	ifp = sc->xn_ifp;
1967 
1968 	if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1969 	if_link_state_change(ifp, LINK_STATE_DOWN);
1970 }
1971 
1972 static void
1973 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1974 {
1975 	int requeue_idx, i;
1976 	grant_ref_t ref;
1977 	netif_rx_request_t *req;
1978 
1979 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1980 		struct mbuf *m;
1981 		u_long pfn;
1982 
1983 		if (rxq->mbufs[i] == NULL)
1984 			continue;
1985 
1986 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1987 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1988 
1989 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1990 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1991 
1992 		gnttab_grant_foreign_access_ref(ref,
1993 		    xenbus_get_otherend_id(rxq->info->xbdev),
1994 		    pfn, 0);
1995 
1996 		req->gref = ref;
1997 		req->id   = requeue_idx;
1998 
1999 		requeue_idx++;
2000 	}
2001 
2002 	rxq->ring.req_prod_pvt = requeue_idx;
2003 }
2004 
2005 /* START of Xenolinux helper functions adapted to FreeBSD */
2006 static int
2007 xn_connect(struct netfront_info *np)
2008 {
2009 	int i, error;
2010 	u_int feature_rx_copy;
2011 	struct netfront_rxq *rxq;
2012 	struct netfront_txq *txq;
2013 
2014 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2015 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
2016 	if (error != 0)
2017 		feature_rx_copy = 0;
2018 
2019 	/* We only support rx copy. */
2020 	if (!feature_rx_copy)
2021 		return (EPROTONOSUPPORT);
2022 
2023 	/* Recovery procedure: */
2024 	error = talk_to_backend(np->xbdev, np);
2025 	if (error != 0)
2026 		return (error);
2027 
2028 	/* Step 1: Reinitialise variables. */
2029 	xn_query_features(np);
2030 	xn_configure_features(np);
2031 
2032 	/* Step 2: Release TX buffer */
2033 	for (i = 0; i < np->num_queues; i++) {
2034 		txq = &np->txq[i];
2035 		xn_release_tx_bufs(txq);
2036 	}
2037 
2038 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
2039 	for (i = 0; i < np->num_queues; i++) {
2040 		rxq = &np->rxq[i];
2041 		xn_rebuild_rx_bufs(rxq);
2042 	}
2043 
2044 	/* Step 4: All public and private state should now be sane.  Get
2045 	 * ready to start sending and receiving packets and give the driver
2046 	 * domain a kick because we've probably just requeued some
2047 	 * packets.
2048 	 */
2049 	netfront_carrier_on(np);
2050 	wakeup(np);
2051 
2052 	return (0);
2053 }
2054 
2055 static void
2056 xn_kick_rings(struct netfront_info *np)
2057 {
2058 	struct netfront_rxq *rxq;
2059 	struct netfront_txq *txq;
2060 	int i;
2061 
2062 	for (i = 0; i < np->num_queues; i++) {
2063 		txq = &np->txq[i];
2064 		rxq = &np->rxq[i];
2065 		xen_intr_signal(txq->xen_intr_handle);
2066 		XN_TX_LOCK(txq);
2067 		xn_txeof(txq);
2068 		XN_TX_UNLOCK(txq);
2069 		XN_RX_LOCK(rxq);
2070 		xn_alloc_rx_buffers(rxq);
2071 		XN_RX_UNLOCK(rxq);
2072 	}
2073 }
2074 
2075 static void
2076 xn_query_features(struct netfront_info *np)
2077 {
2078 	int val;
2079 
2080 	device_printf(np->xbdev, "backend features:");
2081 
2082 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2083 		"feature-sg", NULL, "%d", &val) != 0)
2084 		val = 0;
2085 
2086 	np->maxfrags = 1;
2087 	if (val) {
2088 		np->maxfrags = MAX_TX_REQ_FRAGS;
2089 		printf(" feature-sg");
2090 	}
2091 
2092 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2093 		"feature-gso-tcpv4", NULL, "%d", &val) != 0)
2094 		val = 0;
2095 
2096 	if_setcapabilitiesbit(np->xn_ifp, 0, IFCAP_TSO4 | IFCAP_LRO);
2097 	if (val) {
2098 		if_setcapabilitiesbit(np->xn_ifp, IFCAP_TSO4 | IFCAP_LRO, 0);
2099 		printf(" feature-gso-tcp4");
2100 	}
2101 
2102 	/*
2103 	 * HW CSUM offload is assumed to be available unless
2104 	 * feature-no-csum-offload is set in xenstore.
2105 	 */
2106 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2107 		"feature-no-csum-offload", NULL, "%d", &val) != 0)
2108 		val = 0;
2109 
2110 	if_setcapabilitiesbit(np->xn_ifp, IFCAP_HWCSUM, 0);
2111 	if (val) {
2112 		if_setcapabilitiesbit(np->xn_ifp, 0, IFCAP_HWCSUM);
2113 		printf(" feature-no-csum-offload");
2114 	}
2115 
2116 	printf("\n");
2117 }
2118 
2119 static int
2120 xn_configure_features(struct netfront_info *np)
2121 {
2122 	int err, cap_enabled;
2123 #if (defined(INET) || defined(INET6))
2124 	int i;
2125 #endif
2126 	if_t ifp;
2127 
2128 	ifp = np->xn_ifp;
2129 	err = 0;
2130 
2131 	if ((if_getcapenable(ifp) & if_getcapabilities(ifp)) == if_getcapenable(ifp)) {
2132 		/* Current options are available, no need to do anything. */
2133 		return (0);
2134 	}
2135 
2136 	/* Try to preserve as many options as possible. */
2137 	cap_enabled = if_getcapenable(ifp);
2138 	if_setcapenable(ifp, 0);
2139 	if_sethwassist(ifp, 0);
2140 
2141 #if (defined(INET) || defined(INET6))
2142 	if ((cap_enabled & IFCAP_LRO) != 0)
2143 		for (i = 0; i < np->num_queues; i++)
2144 			tcp_lro_free(&np->rxq[i].lro);
2145 	if (xn_enable_lro &&
2146 	    (if_getcapabilities(ifp) & cap_enabled & IFCAP_LRO) != 0) {
2147 	    	if_setcapenablebit(ifp, IFCAP_LRO, 0);
2148 		for (i = 0; i < np->num_queues; i++) {
2149 			err = tcp_lro_init(&np->rxq[i].lro);
2150 			if (err != 0) {
2151 				device_printf(np->xbdev,
2152 				    "LRO initialization failed\n");
2153 				if_setcapenablebit(ifp, 0, IFCAP_LRO);
2154 				break;
2155 			}
2156 			np->rxq[i].lro.ifp = ifp;
2157 		}
2158 	}
2159 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_TSO4) != 0) {
2160 		if_setcapenablebit(ifp, IFCAP_TSO4, 0);
2161 		if_sethwassistbits(ifp, CSUM_TSO, 0);
2162 	}
2163 #endif
2164 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_TXCSUM) != 0) {
2165 		if_setcapenablebit(ifp, IFCAP_TXCSUM, 0);
2166 		if_sethwassistbits(ifp, XN_CSUM_FEATURES, 0);
2167 	}
2168 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_RXCSUM) != 0)
2169 		if_setcapenablebit(ifp, IFCAP_RXCSUM, 0);
2170 
2171 	return (err);
2172 }
2173 
2174 static int
2175 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2176 {
2177 	struct netfront_info *np;
2178 	if_t ifp;
2179 	struct buf_ring *br;
2180 	int error, notify;
2181 
2182 	np = txq->info;
2183 	br = txq->br;
2184 	ifp = np->xn_ifp;
2185 	error = 0;
2186 
2187 	XN_TX_LOCK_ASSERT(txq);
2188 
2189 	if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 ||
2190 	    !netfront_carrier_ok(np)) {
2191 		if (m != NULL)
2192 			error = drbr_enqueue(ifp, br, m);
2193 		return (error);
2194 	}
2195 
2196 	if (m != NULL) {
2197 		error = drbr_enqueue(ifp, br, m);
2198 		if (error != 0)
2199 			return (error);
2200 	}
2201 
2202 	while ((m = drbr_peek(ifp, br)) != NULL) {
2203 		if (!xn_tx_slot_available(txq)) {
2204 			drbr_putback(ifp, br, m);
2205 			break;
2206 		}
2207 
2208 		error = xn_assemble_tx_request(txq, m);
2209 		/* xn_assemble_tx_request always consumes the mbuf*/
2210 		if (error != 0) {
2211 			drbr_advance(ifp, br);
2212 			break;
2213 		}
2214 
2215 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2216 		if (notify)
2217 			xen_intr_signal(txq->xen_intr_handle);
2218 
2219 		drbr_advance(ifp, br);
2220 	}
2221 
2222 	if (RING_FULL(&txq->ring))
2223 		txq->full = true;
2224 
2225 	return (0);
2226 }
2227 
2228 static int
2229 xn_txq_mq_start(if_t ifp, struct mbuf *m)
2230 {
2231 	struct netfront_info *np;
2232 	struct netfront_txq *txq;
2233 	int i, npairs, error;
2234 
2235 	np = if_getsoftc(ifp);
2236 	npairs = np->num_queues;
2237 
2238 	if (!netfront_carrier_ok(np))
2239 		return (ENOBUFS);
2240 
2241 	KASSERT(npairs != 0, ("called with 0 available queues"));
2242 
2243 	/* check if flowid is set */
2244 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2245 		i = m->m_pkthdr.flowid % npairs;
2246 	else
2247 		i = curcpu % npairs;
2248 
2249 	txq = &np->txq[i];
2250 
2251 	if (XN_TX_TRYLOCK(txq) != 0) {
2252 		error = xn_txq_mq_start_locked(txq, m);
2253 		XN_TX_UNLOCK(txq);
2254 	} else {
2255 		error = drbr_enqueue(ifp, txq->br, m);
2256 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2257 	}
2258 
2259 	return (error);
2260 }
2261 
2262 static void
2263 xn_qflush(if_t ifp)
2264 {
2265 	struct netfront_info *np;
2266 	struct netfront_txq *txq;
2267 	struct mbuf *m;
2268 	int i;
2269 
2270 	np = if_getsoftc(ifp);
2271 
2272 	for (i = 0; i < np->num_queues; i++) {
2273 		txq = &np->txq[i];
2274 
2275 		XN_TX_LOCK(txq);
2276 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2277 			m_freem(m);
2278 		XN_TX_UNLOCK(txq);
2279 	}
2280 
2281 	if_qflush(ifp);
2282 }
2283 
2284 /**
2285  * Create a network device.
2286  * @param dev  Newbus device representing this virtual NIC.
2287  */
2288 int
2289 create_netdev(device_t dev)
2290 {
2291 	struct netfront_info *np;
2292 	int err, cap_enabled;
2293 	if_t ifp;
2294 
2295 	np = device_get_softc(dev);
2296 
2297 	np->xbdev         = dev;
2298 
2299 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2300 
2301 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2302 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2303 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2304 
2305 	err = xen_net_read_mac(dev, np->mac);
2306 	if (err != 0)
2307 		goto error;
2308 
2309 	/* Set up ifnet structure */
2310 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2311 	if_setsoftc(ifp, np);
2312 	if_initname(ifp, "xn",  device_get_unit(dev));
2313 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
2314 	if_setioctlfn(ifp, xn_ioctl);
2315 
2316 	if_settransmitfn(ifp, xn_txq_mq_start);
2317 	if_setqflushfn(ifp, xn_qflush);
2318 
2319 	if_setinitfn(ifp, xn_ifinit);
2320 
2321 	if_sethwassist(ifp, XN_CSUM_FEATURES);
2322 	/* Enable all supported features at device creation. */
2323 	if_setcapabilities(ifp, IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_LRO);
2324 	cap_enabled = if_getcapabilities(ifp);
2325 	if (!xn_enable_lro) {
2326 		cap_enabled &= ~IFCAP_LRO;
2327 	}
2328 	if_setcapenable(ifp, cap_enabled);
2329 
2330 	if_sethwtsomax(ifp, 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN));
2331 	if_sethwtsomaxsegcount(ifp, MAX_TX_REQ_FRAGS);
2332 	if_sethwtsomaxsegsize(ifp, PAGE_SIZE);
2333 
2334 	ether_ifattach(ifp, np->mac);
2335 	netfront_carrier_off(np);
2336 
2337 	err = bus_dma_tag_create(
2338 	    bus_get_dma_tag(dev),		/* parent */
2339 	    1, PAGE_SIZE,			/* algnmnt, boundary */
2340 	    BUS_SPACE_MAXADDR,			/* lowaddr */
2341 	    BUS_SPACE_MAXADDR,			/* highaddr */
2342 	    NULL, NULL,				/* filter, filterarg */
2343 	    PAGE_SIZE * MAX_TX_REQ_FRAGS,	/* max request size */
2344 	    MAX_TX_REQ_FRAGS,			/* max segments */
2345 	    PAGE_SIZE,				/* maxsegsize */
2346 	    BUS_DMA_ALLOCNOW,			/* flags */
2347 	    NULL, NULL,				/* lockfunc, lockarg */
2348 	    &np->dma_tag);
2349 
2350 	return (err);
2351 
2352 error:
2353 	KASSERT(err != 0, ("Error path with no error code specified"));
2354 	return (err);
2355 }
2356 
2357 static int
2358 netfront_detach(device_t dev)
2359 {
2360 	struct netfront_info *info = device_get_softc(dev);
2361 
2362 	DPRINTK("%s\n", xenbus_get_node(dev));
2363 
2364 	netif_free(info);
2365 
2366 	return 0;
2367 }
2368 
2369 static void
2370 netif_free(struct netfront_info *np)
2371 {
2372 
2373 	XN_LOCK(np);
2374 	xn_stop(np);
2375 	XN_UNLOCK(np);
2376 	netif_disconnect_backend(np);
2377 	ether_ifdetach(np->xn_ifp);
2378 	free(np->rxq, M_DEVBUF);
2379 	free(np->txq, M_DEVBUF);
2380 	if_free(np->xn_ifp);
2381 	np->xn_ifp = NULL;
2382 	ifmedia_removeall(&np->sc_media);
2383 	bus_dma_tag_destroy(np->dma_tag);
2384 }
2385 
2386 static void
2387 netif_disconnect_backend(struct netfront_info *np)
2388 {
2389 	u_int i;
2390 
2391 	for (i = 0; i < np->num_queues; i++) {
2392 		XN_RX_LOCK(&np->rxq[i]);
2393 		XN_TX_LOCK(&np->txq[i]);
2394 	}
2395 	netfront_carrier_off(np);
2396 	for (i = 0; i < np->num_queues; i++) {
2397 		XN_RX_UNLOCK(&np->rxq[i]);
2398 		XN_TX_UNLOCK(&np->txq[i]);
2399 	}
2400 
2401 	for (i = 0; i < np->num_queues; i++) {
2402 		disconnect_rxq(&np->rxq[i]);
2403 		disconnect_txq(&np->txq[i]);
2404 	}
2405 }
2406 
2407 static int
2408 xn_ifmedia_upd(if_t ifp)
2409 {
2410 
2411 	return (0);
2412 }
2413 
2414 static void
2415 xn_ifmedia_sts(if_t ifp, struct ifmediareq *ifmr)
2416 {
2417 
2418 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2419 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2420 }
2421 
2422 /* ** Driver registration ** */
2423 static device_method_t netfront_methods[] = {
2424 	/* Device interface */
2425 	DEVMETHOD(device_probe,         netfront_probe),
2426 	DEVMETHOD(device_attach,        netfront_attach),
2427 	DEVMETHOD(device_detach,        netfront_detach),
2428 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2429 	DEVMETHOD(device_suspend,       netfront_suspend),
2430 	DEVMETHOD(device_resume,        netfront_resume),
2431 
2432 	/* Xenbus interface */
2433 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2434 
2435 	DEVMETHOD_END
2436 };
2437 
2438 static driver_t netfront_driver = {
2439 	"xn",
2440 	netfront_methods,
2441 	sizeof(struct netfront_info),
2442 };
2443 
2444 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, NULL, NULL);
2445