xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision 02f3b17fa5967b15d3841d9f7df56264557d1bfb)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 
36 #include <sys/param.h>
37 #include <sys/sockio.h>
38 #include <sys/limits.h>
39 #include <sys/mbuf.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/kernel.h>
43 #include <sys/socket.h>
44 #include <sys/sysctl.h>
45 #include <sys/taskqueue.h>
46 
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/if_arp.h>
50 #include <net/ethernet.h>
51 #include <net/if_media.h>
52 #include <net/bpf.h>
53 #include <net/if_types.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/if_ether.h>
58 #include <netinet/tcp.h>
59 #include <netinet/tcp_lro.h>
60 
61 #include <vm/vm.h>
62 #include <vm/pmap.h>
63 
64 #include <sys/bus.h>
65 
66 #include <xen/xen-os.h>
67 #include <xen/hypervisor.h>
68 #include <xen/xen_intr.h>
69 #include <xen/gnttab.h>
70 #include <contrib/xen/memory.h>
71 #include <contrib/xen/io/netif.h>
72 #include <xen/xenbus/xenbusvar.h>
73 
74 #include <machine/bus.h>
75 
76 #include "xenbus_if.h"
77 
78 /* Features supported by all backends.  TSO and LRO can be negotiated */
79 #define XN_CSUM_FEATURES	(CSUM_TCP | CSUM_UDP)
80 
81 #define NET_TX_RING_SIZE __CONST_RING_SIZE(netif_tx, PAGE_SIZE)
82 #define NET_RX_RING_SIZE __CONST_RING_SIZE(netif_rx, PAGE_SIZE)
83 
84 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
85 
86 /*
87  * Should the driver do LRO on the RX end
88  *  this can be toggled on the fly, but the
89  *  interface must be reset (down/up) for it
90  *  to take effect.
91  */
92 static int xn_enable_lro = 1;
93 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
94 
95 /*
96  * Number of pairs of queues.
97  */
98 static unsigned long xn_num_queues = 4;
99 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
100 
101 /**
102  * \brief The maximum allowed data fragments in a single transmit
103  *        request.
104  *
105  * This limit is imposed by the backend driver.  We assume here that
106  * we are dealing with a Linux driver domain and have set our limit
107  * to mirror the Linux MAX_SKB_FRAGS constant.
108  */
109 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
110 
111 #define RX_COPY_THRESHOLD 256
112 
113 #define net_ratelimit() 0
114 
115 struct netfront_rxq;
116 struct netfront_txq;
117 struct netfront_info;
118 struct netfront_rx_info;
119 
120 static void xn_txeof(struct netfront_txq *);
121 static void xn_rxeof(struct netfront_rxq *);
122 static void xn_alloc_rx_buffers(struct netfront_rxq *);
123 static void xn_alloc_rx_buffers_callout(void *arg);
124 
125 static void xn_release_rx_bufs(struct netfront_rxq *);
126 static void xn_release_tx_bufs(struct netfront_txq *);
127 
128 static void xn_rxq_intr(struct netfront_rxq *);
129 static void xn_txq_intr(struct netfront_txq *);
130 static void xn_intr(void *);
131 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
132 static int xn_ioctl(if_t, u_long, caddr_t);
133 static void xn_ifinit_locked(struct netfront_info *);
134 static void xn_ifinit(void *);
135 static void xn_stop(struct netfront_info *);
136 static void xn_query_features(struct netfront_info *np);
137 static int xn_configure_features(struct netfront_info *np);
138 static void netif_free(struct netfront_info *info);
139 static int netfront_detach(device_t dev);
140 
141 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
142 static int xn_txq_mq_start(if_t, struct mbuf *);
143 
144 static int talk_to_backend(device_t dev, struct netfront_info *info);
145 static int create_netdev(device_t dev);
146 static void netif_disconnect_backend(struct netfront_info *info);
147 static int setup_device(device_t dev, struct netfront_info *info,
148     unsigned long);
149 static int xn_ifmedia_upd(if_t ifp);
150 static void xn_ifmedia_sts(if_t ifp, struct ifmediareq *ifmr);
151 
152 static int xn_connect(struct netfront_info *);
153 static void xn_kick_rings(struct netfront_info *);
154 
155 static int xn_get_responses(struct netfront_rxq *,
156     struct netfront_rx_info *, RING_IDX, RING_IDX *,
157     struct mbuf **);
158 
159 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
160 
161 #define INVALID_P2M_ENTRY (~0UL)
162 #define XN_QUEUE_NAME_LEN  8	/* xn{t,r}x_%u, allow for two digits */
163 struct netfront_rxq {
164 	struct netfront_info 	*info;
165 	u_int			id;
166 	char			name[XN_QUEUE_NAME_LEN];
167 	struct mtx		lock;
168 
169 	int			ring_ref;
170 	netif_rx_front_ring_t 	ring;
171 	xen_intr_handle_t	xen_intr_handle;
172 
173 	grant_ref_t 		gref_head;
174 	grant_ref_t 		grant_ref[NET_RX_RING_SIZE + 1];
175 
176 	struct mbuf		*mbufs[NET_RX_RING_SIZE + 1];
177 
178 	struct lro_ctrl		lro;
179 
180 	struct callout		rx_refill;
181 };
182 
183 struct netfront_txq {
184 	struct netfront_info 	*info;
185 	u_int 			id;
186 	char			name[XN_QUEUE_NAME_LEN];
187 	struct mtx		lock;
188 
189 	int			ring_ref;
190 	netif_tx_front_ring_t	ring;
191 	xen_intr_handle_t 	xen_intr_handle;
192 
193 	grant_ref_t		gref_head;
194 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
195 
196 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
197 	int			mbufs_cnt;
198 	struct buf_ring		*br;
199 
200 	struct taskqueue 	*tq;
201 	struct task       	defrtask;
202 
203 	bus_dma_segment_t	segs[MAX_TX_REQ_FRAGS];
204 	struct mbuf_xennet {
205 		struct m_tag 	tag;
206 		bus_dma_tag_t	dma_tag;
207 		bus_dmamap_t	dma_map;
208 		struct netfront_txq *txq;
209 		SLIST_ENTRY(mbuf_xennet) next;
210 		u_int 		count;
211 	}			xennet_tag[NET_TX_RING_SIZE + 1];
212 	SLIST_HEAD(, mbuf_xennet) tags;
213 
214 	bool			full;
215 };
216 
217 struct netfront_info {
218 	if_t			xn_ifp;
219 
220 	struct mtx   		sc_lock;
221 
222 	u_int  num_queues;
223 	struct netfront_rxq 	*rxq;
224 	struct netfront_txq 	*txq;
225 
226 	u_int			carrier;
227 	u_int			maxfrags;
228 
229 	device_t		xbdev;
230 	uint8_t			mac[ETHER_ADDR_LEN];
231 
232 	int			xn_if_flags;
233 
234 	struct ifmedia		sc_media;
235 
236 	bus_dma_tag_t		dma_tag;
237 
238 	bool			xn_reset;
239 };
240 
241 struct netfront_rx_info {
242 	struct netif_rx_response rx;
243 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
244 };
245 
246 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
247 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
248 
249 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
250 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
251 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
252 
253 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
254 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
255 
256 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
257 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
258 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
259 
260 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
261 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
262 #define netfront_carrier_ok(netif)	((netif)->carrier)
263 
264 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
265 
266 static inline void
267 add_id_to_freelist(struct mbuf **list, uintptr_t id)
268 {
269 
270 	KASSERT(id != 0,
271 		("%s: the head item (0) must always be free.", __func__));
272 	list[id] = list[0];
273 	list[0]  = (struct mbuf *)id;
274 }
275 
276 static inline unsigned short
277 get_id_from_freelist(struct mbuf **list)
278 {
279 	uintptr_t id;
280 
281 	id = (uintptr_t)list[0];
282 	KASSERT(id != 0,
283 		("%s: the head item (0) must always remain free.", __func__));
284 	list[0] = list[id];
285 	return (id);
286 }
287 
288 static inline int
289 xn_rxidx(RING_IDX idx)
290 {
291 
292 	return idx & (NET_RX_RING_SIZE - 1);
293 }
294 
295 static inline struct mbuf *
296 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
297 {
298 	int i;
299 	struct mbuf *m;
300 
301 	i = xn_rxidx(ri);
302 	m = rxq->mbufs[i];
303 	rxq->mbufs[i] = NULL;
304 	return (m);
305 }
306 
307 static inline grant_ref_t
308 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
309 {
310 	int i = xn_rxidx(ri);
311 	grant_ref_t ref = rxq->grant_ref[i];
312 
313 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
314 	rxq->grant_ref[i] = GRANT_REF_INVALID;
315 	return (ref);
316 }
317 
318 #define MTAG_COOKIE 1218492000
319 #define MTAG_XENNET 0
320 
321 static void mbuf_grab(struct mbuf *m)
322 {
323 	struct mbuf_xennet *ref;
324 
325 	ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
326 	    MTAG_XENNET, NULL);
327 	KASSERT(ref != NULL, ("Cannot find refcount"));
328 	ref->count++;
329 }
330 
331 static void mbuf_release(struct mbuf *m)
332 {
333 	struct mbuf_xennet *ref;
334 
335 	ref = (struct mbuf_xennet *)m_tag_locate(m, MTAG_COOKIE,
336 	    MTAG_XENNET, NULL);
337 	KASSERT(ref != NULL, ("Cannot find refcount"));
338 	KASSERT(ref->count > 0, ("Invalid reference count"));
339 
340 	if (--ref->count == 0)
341 		m_freem(m);
342 }
343 
344 static void tag_free(struct m_tag *t)
345 {
346 	struct mbuf_xennet *ref = (struct mbuf_xennet *)t;
347 
348 	KASSERT(ref->count == 0, ("Free mbuf tag with pending refcnt"));
349 	bus_dmamap_sync(ref->dma_tag, ref->dma_map, BUS_DMASYNC_POSTWRITE);
350 	bus_dmamap_destroy(ref->dma_tag, ref->dma_map);
351 	SLIST_INSERT_HEAD(&ref->txq->tags, ref, next);
352 }
353 
354 #define IPRINTK(fmt, args...) \
355     printf("[XEN] " fmt, ##args)
356 #ifdef INVARIANTS
357 #define WPRINTK(fmt, args...) \
358     printf("[XEN] " fmt, ##args)
359 #else
360 #define WPRINTK(fmt, args...)
361 #endif
362 #ifdef DEBUG
363 #define DPRINTK(fmt, args...) \
364     printf("[XEN] %s: " fmt, __func__, ##args)
365 #else
366 #define DPRINTK(fmt, args...)
367 #endif
368 
369 /**
370  * Read the 'mac' node at the given device's node in the store, and parse that
371  * as colon-separated octets, placing result the given mac array.  mac must be
372  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
373  * Return 0 on success, or errno on error.
374  */
375 static int
376 xen_net_read_mac(device_t dev, uint8_t mac[])
377 {
378 	int error, i;
379 	char *s, *e, *macstr;
380 	const char *path;
381 
382 	path = xenbus_get_node(dev);
383 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
384 	if (error == ENOENT) {
385 		/*
386 		 * Deal with missing mac XenStore nodes on devices with
387 		 * HVM emulation (the 'ioemu' configuration attribute)
388 		 * enabled.
389 		 *
390 		 * The HVM emulator may execute in a stub device model
391 		 * domain which lacks the permission, only given to Dom0,
392 		 * to update the guest's XenStore tree.  For this reason,
393 		 * the HVM emulator doesn't even attempt to write the
394 		 * front-side mac node, even when operating in Dom0.
395 		 * However, there should always be a mac listed in the
396 		 * backend tree.  Fallback to this version if our query
397 		 * of the front side XenStore location doesn't find
398 		 * anything.
399 		 */
400 		path = xenbus_get_otherend_path(dev);
401 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
402 	}
403 	if (error != 0) {
404 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
405 		return (error);
406 	}
407 
408 	s = macstr;
409 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
410 		mac[i] = strtoul(s, &e, 16);
411 		if (s == e || (e[0] != ':' && e[0] != 0)) {
412 			free(macstr, M_XENBUS);
413 			return (ENOENT);
414 		}
415 		s = &e[1];
416 	}
417 	free(macstr, M_XENBUS);
418 	return (0);
419 }
420 
421 /**
422  * Entry point to this code when a new device is created.  Allocate the basic
423  * structures and the ring buffers for communication with the backend, and
424  * inform the backend of the appropriate details for those.  Switch to
425  * Connected state.
426  */
427 static int
428 netfront_probe(device_t dev)
429 {
430 
431 	if (xen_pv_nics_disabled())
432 		return (ENXIO);
433 
434 	if (!strcmp(xenbus_get_type(dev), "vif")) {
435 		device_set_desc(dev, "Virtual Network Interface");
436 		return (0);
437 	}
438 
439 	return (ENXIO);
440 }
441 
442 static int
443 netfront_attach(device_t dev)
444 {
445 	int err;
446 
447 	err = create_netdev(dev);
448 	if (err != 0) {
449 		xenbus_dev_fatal(dev, err, "creating netdev");
450 		return (err);
451 	}
452 
453 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
454 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
455 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
456 	    &xn_enable_lro, 0, "Large Receive Offload");
457 
458 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
459 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
460 	    OID_AUTO, "num_queues", CTLFLAG_RD,
461 	    &xn_num_queues, "Number of pairs of queues");
462 
463 	return (0);
464 }
465 
466 static int
467 netfront_suspend(device_t dev)
468 {
469 	struct netfront_info *np = device_get_softc(dev);
470 	u_int i;
471 
472 	for (i = 0; i < np->num_queues; i++) {
473 		XN_RX_LOCK(&np->rxq[i]);
474 		XN_TX_LOCK(&np->txq[i]);
475 	}
476 	netfront_carrier_off(np);
477 	for (i = 0; i < np->num_queues; i++) {
478 		XN_RX_UNLOCK(&np->rxq[i]);
479 		XN_TX_UNLOCK(&np->txq[i]);
480 	}
481 	return (0);
482 }
483 
484 /**
485  * We are reconnecting to the backend, due to a suspend/resume, or a backend
486  * driver restart.  We tear down our netif structure and recreate it, but
487  * leave the device-layer structures intact so that this is transparent to the
488  * rest of the kernel.
489  */
490 static int
491 netfront_resume(device_t dev)
492 {
493 	struct netfront_info *info = device_get_softc(dev);
494 	u_int i;
495 
496 	if (xen_suspend_cancelled) {
497 		for (i = 0; i < info->num_queues; i++) {
498 			XN_RX_LOCK(&info->rxq[i]);
499 			XN_TX_LOCK(&info->txq[i]);
500 		}
501 		netfront_carrier_on(info);
502 		for (i = 0; i < info->num_queues; i++) {
503 			XN_RX_UNLOCK(&info->rxq[i]);
504 			XN_TX_UNLOCK(&info->txq[i]);
505 		}
506 		return (0);
507 	}
508 
509 	netif_disconnect_backend(info);
510 	return (0);
511 }
512 
513 static int
514 write_queue_xenstore_keys(device_t dev,
515     struct netfront_rxq *rxq,
516     struct netfront_txq *txq,
517     struct xs_transaction *xst, bool hierarchy)
518 {
519 	int err;
520 	const char *message;
521 	const char *node = xenbus_get_node(dev);
522 	char *path;
523 	size_t path_size;
524 
525 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
526 	/* Split event channel support is not yet there. */
527 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
528 	    ("Split event channels are not supported"));
529 
530 	if (hierarchy) {
531 		path_size = strlen(node) + 10;
532 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
533 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
534 	} else {
535 		path_size = strlen(node) + 1;
536 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
537 		snprintf(path, path_size, "%s", node);
538 	}
539 
540 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
541 	if (err != 0) {
542 		message = "writing tx ring-ref";
543 		goto error;
544 	}
545 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
546 	if (err != 0) {
547 		message = "writing rx ring-ref";
548 		goto error;
549 	}
550 	err = xs_printf(*xst, path, "event-channel", "%u",
551 	    xen_intr_port(rxq->xen_intr_handle));
552 	if (err != 0) {
553 		message = "writing event-channel";
554 		goto error;
555 	}
556 
557 	free(path, M_DEVBUF);
558 
559 	return (0);
560 
561 error:
562 	free(path, M_DEVBUF);
563 	xenbus_dev_fatal(dev, err, "%s", message);
564 
565 	return (err);
566 }
567 
568 /* Common code used when first setting up, and when resuming. */
569 static int
570 talk_to_backend(device_t dev, struct netfront_info *info)
571 {
572 	const char *message;
573 	struct xs_transaction xst;
574 	const char *node = xenbus_get_node(dev);
575 	int err;
576 	unsigned long num_queues, max_queues = 0;
577 	unsigned int i;
578 
579 	err = xen_net_read_mac(dev, info->mac);
580 	if (err != 0) {
581 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
582 		goto out;
583 	}
584 
585 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
586 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
587 	if (err != 0)
588 		max_queues = 1;
589 	num_queues = xn_num_queues;
590 	if (num_queues > max_queues)
591 		num_queues = max_queues;
592 
593 	err = setup_device(dev, info, num_queues);
594 	if (err != 0)
595 		goto out;
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 	gnttab_end_foreign_access(rxq->ring_ref, NULL);
723 	/*
724 	 * No split event channel support at the moment, handle will
725 	 * be unbound in tx. So no need to call xen_intr_unbind here,
726 	 * but we do want to reset the handler to 0.
727 	 */
728 	rxq->xen_intr_handle = 0;
729 }
730 
731 static void
732 destroy_rxq(struct netfront_rxq *rxq)
733 {
734 
735 	callout_drain(&rxq->rx_refill);
736 	free(rxq->ring.sring, M_DEVBUF);
737 }
738 
739 static void
740 destroy_rxqs(struct netfront_info *np)
741 {
742 	int i;
743 
744 	for (i = 0; i < np->num_queues; i++)
745 		destroy_rxq(&np->rxq[i]);
746 
747 	free(np->rxq, M_DEVBUF);
748 	np->rxq = NULL;
749 }
750 
751 static int
752 setup_rxqs(device_t dev, struct netfront_info *info,
753 	   unsigned long num_queues)
754 {
755 	int q, i;
756 	int error;
757 	netif_rx_sring_t *rxs;
758 	struct netfront_rxq *rxq;
759 
760 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
761 	    M_DEVBUF, M_WAITOK|M_ZERO);
762 
763 	for (q = 0; q < num_queues; q++) {
764 		rxq = &info->rxq[q];
765 
766 		rxq->id = q;
767 		rxq->info = info;
768 		rxq->ring_ref = GRANT_REF_INVALID;
769 		rxq->ring.sring = NULL;
770 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
771 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
772 		    MTX_DEF);
773 
774 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
775 			rxq->mbufs[i] = NULL;
776 			rxq->grant_ref[i] = GRANT_REF_INVALID;
777 		}
778 
779 		/* Start resources allocation */
780 
781 		if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
782 		    &rxq->gref_head) != 0) {
783 			device_printf(dev, "allocating rx gref");
784 			error = ENOMEM;
785 			goto fail;
786 		}
787 
788 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
789 		    M_WAITOK|M_ZERO);
790 		SHARED_RING_INIT(rxs);
791 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
792 
793 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
794 		    &rxq->ring_ref);
795 		if (error != 0) {
796 			device_printf(dev, "granting rx ring page");
797 			goto fail_grant_ring;
798 		}
799 
800 		callout_init(&rxq->rx_refill, 1);
801 	}
802 
803 	return (0);
804 
805 fail_grant_ring:
806 	gnttab_free_grant_references(rxq->gref_head);
807 	free(rxq->ring.sring, M_DEVBUF);
808 fail:
809 	for (; q >= 0; q--) {
810 		disconnect_rxq(&info->rxq[q]);
811 		destroy_rxq(&info->rxq[q]);
812 	}
813 
814 	free(info->rxq, M_DEVBUF);
815 	return (error);
816 }
817 
818 static void
819 disconnect_txq(struct netfront_txq *txq)
820 {
821 
822 	xn_release_tx_bufs(txq);
823 	gnttab_free_grant_references(txq->gref_head);
824 	gnttab_end_foreign_access(txq->ring_ref, NULL);
825 	xen_intr_unbind(&txq->xen_intr_handle);
826 }
827 
828 static void
829 destroy_txq(struct netfront_txq *txq)
830 {
831 	unsigned int i;
832 
833 	free(txq->ring.sring, M_DEVBUF);
834 	buf_ring_free(txq->br, M_DEVBUF);
835 	taskqueue_drain_all(txq->tq);
836 	taskqueue_free(txq->tq);
837 
838 	for (i = 0; i <= NET_TX_RING_SIZE; i++) {
839 		bus_dmamap_destroy(txq->info->dma_tag,
840 		    txq->xennet_tag[i].dma_map);
841 		txq->xennet_tag[i].dma_map = NULL;
842 	}
843 }
844 
845 static void
846 destroy_txqs(struct netfront_info *np)
847 {
848 	int i;
849 
850 	for (i = 0; i < np->num_queues; i++)
851 		destroy_txq(&np->txq[i]);
852 
853 	free(np->txq, M_DEVBUF);
854 	np->txq = NULL;
855 }
856 
857 static int
858 setup_txqs(device_t dev, struct netfront_info *info,
859 	   unsigned long num_queues)
860 {
861 	int q, i;
862 	int error;
863 	netif_tx_sring_t *txs;
864 	struct netfront_txq *txq;
865 
866 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
867 	    M_DEVBUF, M_WAITOK|M_ZERO);
868 
869 	for (q = 0; q < num_queues; q++) {
870 		txq = &info->txq[q];
871 
872 		txq->id = q;
873 		txq->info = info;
874 
875 		txq->ring_ref = GRANT_REF_INVALID;
876 		txq->ring.sring = NULL;
877 
878 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
879 
880 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
881 		    MTX_DEF);
882 		SLIST_INIT(&txq->tags);
883 
884 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
885 			txq->mbufs[i] = (void *) ((u_long) i+1);
886 			txq->grant_ref[i] = GRANT_REF_INVALID;
887 			txq->xennet_tag[i].txq = txq;
888 			txq->xennet_tag[i].dma_tag = info->dma_tag;
889 			error = bus_dmamap_create(info->dma_tag, 0,
890 			    &txq->xennet_tag[i].dma_map);
891 			if (error != 0) {
892 				device_printf(dev,
893 				    "failed to allocate dma map\n");
894 				goto fail;
895 			}
896 			m_tag_setup(&txq->xennet_tag[i].tag,
897 			    MTAG_COOKIE, MTAG_XENNET,
898 			    sizeof(txq->xennet_tag[i]) -
899 			    sizeof(txq->xennet_tag[i].tag));
900 			txq->xennet_tag[i].tag.m_tag_free = &tag_free;
901 			SLIST_INSERT_HEAD(&txq->tags, &txq->xennet_tag[i],
902 			    next);
903 		}
904 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
905 
906 		/* Start resources allocation. */
907 
908 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
909 		    &txq->gref_head) != 0) {
910 			device_printf(dev, "failed to allocate tx grant refs\n");
911 			error = ENOMEM;
912 			goto fail;
913 		}
914 
915 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
916 		    M_WAITOK|M_ZERO);
917 		SHARED_RING_INIT(txs);
918 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
919 
920 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
921 		    &txq->ring_ref);
922 		if (error != 0) {
923 			device_printf(dev, "failed to grant tx ring\n");
924 			goto fail_grant_ring;
925 		}
926 
927 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
928 		    M_WAITOK, &txq->lock);
929 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
930 
931 		txq->tq = taskqueue_create(txq->name, M_WAITOK,
932 		    taskqueue_thread_enqueue, &txq->tq);
933 
934 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
935 		    "%s txq %d", device_get_nameunit(dev), txq->id);
936 		if (error != 0) {
937 			device_printf(dev, "failed to start tx taskq %d\n",
938 			    txq->id);
939 			goto fail_start_thread;
940 		}
941 
942 		error = xen_intr_alloc_and_bind_local_port(dev,
943 		    xenbus_get_otherend_id(dev), /* filter */ NULL, xn_intr,
944 		    &info->txq[q], INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
945 		    &txq->xen_intr_handle);
946 
947 		if (error != 0) {
948 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
949 			goto fail_bind_port;
950 		}
951 	}
952 
953 	return (0);
954 
955 fail_bind_port:
956 	taskqueue_drain_all(txq->tq);
957 fail_start_thread:
958 	buf_ring_free(txq->br, M_DEVBUF);
959 	taskqueue_free(txq->tq);
960 	gnttab_end_foreign_access(txq->ring_ref, NULL);
961 fail_grant_ring:
962 	gnttab_free_grant_references(txq->gref_head);
963 	free(txq->ring.sring, M_DEVBUF);
964 fail:
965 	for (; q >= 0; q--) {
966 		disconnect_txq(&info->txq[q]);
967 		destroy_txq(&info->txq[q]);
968 	}
969 
970 	free(info->txq, M_DEVBUF);
971 	return (error);
972 }
973 
974 static int
975 setup_device(device_t dev, struct netfront_info *info,
976     unsigned long num_queues)
977 {
978 	int error;
979 	int q;
980 
981 	if (info->txq)
982 		destroy_txqs(info);
983 
984 	if (info->rxq)
985 		destroy_rxqs(info);
986 
987 	info->num_queues = 0;
988 
989 	error = setup_rxqs(dev, info, num_queues);
990 	if (error != 0)
991 		goto out;
992 	error = setup_txqs(dev, info, num_queues);
993 	if (error != 0)
994 		goto out;
995 
996 	info->num_queues = num_queues;
997 
998 	/* No split event channel at the moment. */
999 	for (q = 0; q < num_queues; q++)
1000 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
1001 
1002 	return (0);
1003 
1004 out:
1005 	KASSERT(error != 0, ("Error path taken without providing an error code"));
1006 	return (error);
1007 }
1008 
1009 #ifdef INET
1010 static u_int
1011 netfront_addr_cb(void *arg, struct ifaddr *a, u_int count)
1012 {
1013 	arp_ifinit((if_t)arg, a);
1014 	return (1);
1015 }
1016 /**
1017  * If this interface has an ipv4 address, send an arp for it. This
1018  * helps to get the network going again after migrating hosts.
1019  */
1020 static void
1021 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
1022 {
1023 	if_t ifp;
1024 
1025 	ifp = info->xn_ifp;
1026 	if_foreach_addr_type(ifp, AF_INET, netfront_addr_cb, ifp);
1027 }
1028 #endif
1029 
1030 /**
1031  * Callback received when the backend's state changes.
1032  */
1033 static void
1034 netfront_backend_changed(device_t dev, XenbusState newstate)
1035 {
1036 	struct netfront_info *sc = device_get_softc(dev);
1037 
1038 	DPRINTK("newstate=%d\n", newstate);
1039 
1040 	CURVNET_SET(if_getvnet(sc->xn_ifp));
1041 
1042 	switch (newstate) {
1043 	case XenbusStateInitialising:
1044 	case XenbusStateInitialised:
1045 	case XenbusStateUnknown:
1046 	case XenbusStateReconfigured:
1047 	case XenbusStateReconfiguring:
1048 		break;
1049 	case XenbusStateInitWait:
1050 		if (xenbus_get_state(dev) != XenbusStateInitialising)
1051 			break;
1052 		if (xn_connect(sc) != 0)
1053 			break;
1054 		/* Switch to connected state before kicking the rings. */
1055 		xenbus_set_state(sc->xbdev, XenbusStateConnected);
1056 		xn_kick_rings(sc);
1057 		break;
1058 	case XenbusStateClosing:
1059 		xenbus_set_state(dev, XenbusStateClosed);
1060 		break;
1061 	case XenbusStateClosed:
1062 		if (sc->xn_reset) {
1063 			netif_disconnect_backend(sc);
1064 			xenbus_set_state(dev, XenbusStateInitialising);
1065 			sc->xn_reset = false;
1066 		}
1067 		break;
1068 	case XenbusStateConnected:
1069 #ifdef INET
1070 		netfront_send_fake_arp(dev, sc);
1071 #endif
1072 		break;
1073 	}
1074 
1075 	CURVNET_RESTORE();
1076 }
1077 
1078 /**
1079  * \brief Verify that there is sufficient space in the Tx ring
1080  *        buffer for a maximally sized request to be enqueued.
1081  *
1082  * A transmit request requires a transmit descriptor for each packet
1083  * fragment, plus up to 2 entries for "options" (e.g. TSO).
1084  */
1085 static inline int
1086 xn_tx_slot_available(struct netfront_txq *txq)
1087 {
1088 
1089 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1090 }
1091 
1092 static void
1093 xn_release_tx_bufs(struct netfront_txq *txq)
1094 {
1095 	int i;
1096 
1097 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1098 		struct mbuf *m;
1099 
1100 		m = txq->mbufs[i];
1101 
1102 		/*
1103 		 * We assume that no kernel addresses are
1104 		 * less than NET_TX_RING_SIZE.  Any entry
1105 		 * in the table that is below this number
1106 		 * must be an index from free-list tracking.
1107 		 */
1108 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1109 			continue;
1110 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1111 		gnttab_release_grant_reference(&txq->gref_head,
1112 		    txq->grant_ref[i]);
1113 		txq->grant_ref[i] = GRANT_REF_INVALID;
1114 		add_id_to_freelist(txq->mbufs, i);
1115 		txq->mbufs_cnt--;
1116 		if (txq->mbufs_cnt < 0) {
1117 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1118 		}
1119 		mbuf_release(m);
1120 	}
1121 }
1122 
1123 static struct mbuf *
1124 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1125 {
1126 	struct mbuf *m;
1127 
1128 	m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1129 	if (m == NULL)
1130 		return NULL;
1131 	m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1132 
1133 	return (m);
1134 }
1135 
1136 static void
1137 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1138 {
1139 	RING_IDX req_prod;
1140 	int notify;
1141 
1142 	XN_RX_LOCK_ASSERT(rxq);
1143 
1144 	if (__predict_false(rxq->info->carrier == 0))
1145 		return;
1146 
1147 	for (req_prod = rxq->ring.req_prod_pvt;
1148 	     req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1149 	     req_prod++) {
1150 		struct mbuf *m;
1151 		unsigned short id;
1152 		grant_ref_t ref;
1153 		struct netif_rx_request *req;
1154 		unsigned long pfn;
1155 
1156 		m = xn_alloc_one_rx_buffer(rxq);
1157 		if (m == NULL)
1158 			break;
1159 
1160 		id = xn_rxidx(req_prod);
1161 
1162 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1163 		rxq->mbufs[id] = m;
1164 
1165 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1166 		KASSERT(ref != GNTTAB_LIST_END,
1167 		    ("reserved grant references exhuasted"));
1168 		rxq->grant_ref[id] = ref;
1169 
1170 		pfn = atop(vtophys(mtod(m, vm_offset_t)));
1171 		req = RING_GET_REQUEST(&rxq->ring, req_prod);
1172 
1173 		gnttab_grant_foreign_access_ref(ref,
1174 		    xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1175 		req->id = id;
1176 		req->gref = ref;
1177 	}
1178 
1179 	rxq->ring.req_prod_pvt = req_prod;
1180 
1181 	/* Not enough requests? Try again later. */
1182 	if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1183 		callout_reset_curcpu(&rxq->rx_refill, hz/10,
1184 		    xn_alloc_rx_buffers_callout, rxq);
1185 		return;
1186 	}
1187 
1188 	wmb();		/* barrier so backend seens requests */
1189 
1190 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1191 	if (notify)
1192 		xen_intr_signal(rxq->xen_intr_handle);
1193 }
1194 
1195 static void xn_alloc_rx_buffers_callout(void *arg)
1196 {
1197 	struct netfront_rxq *rxq;
1198 
1199 	rxq = (struct netfront_rxq *)arg;
1200 	XN_RX_LOCK(rxq);
1201 	xn_alloc_rx_buffers(rxq);
1202 	XN_RX_UNLOCK(rxq);
1203 }
1204 
1205 static void
1206 xn_release_rx_bufs(struct netfront_rxq *rxq)
1207 {
1208 	int i,  ref;
1209 	struct mbuf *m;
1210 
1211 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1212 		m = rxq->mbufs[i];
1213 
1214 		if (m == NULL)
1215 			continue;
1216 
1217 		ref = rxq->grant_ref[i];
1218 		if (ref == GRANT_REF_INVALID)
1219 			continue;
1220 
1221 		gnttab_end_foreign_access_ref(ref);
1222 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1223 		rxq->mbufs[i] = NULL;
1224 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1225 		m_freem(m);
1226 	}
1227 }
1228 
1229 static void
1230 xn_rxeof(struct netfront_rxq *rxq)
1231 {
1232 	if_t ifp;
1233 	struct netfront_info *np = rxq->info;
1234 #if (defined(INET) || defined(INET6))
1235 	struct lro_ctrl *lro = &rxq->lro;
1236 #endif
1237 	struct netfront_rx_info rinfo;
1238 	struct netif_rx_response *rx = &rinfo.rx;
1239 	struct netif_extra_info *extras = rinfo.extras;
1240 	RING_IDX i, rp;
1241 	struct mbuf *m;
1242 	struct mbufq mbufq_rxq, mbufq_errq;
1243 	int err, work_to_do;
1244 
1245 	XN_RX_LOCK_ASSERT(rxq);
1246 
1247 	if (!netfront_carrier_ok(np))
1248 		return;
1249 
1250 	/* XXX: there should be some sane limit. */
1251 	mbufq_init(&mbufq_errq, INT_MAX);
1252 	mbufq_init(&mbufq_rxq, INT_MAX);
1253 
1254 	ifp = np->xn_ifp;
1255 
1256 	do {
1257 		rp = rxq->ring.sring->rsp_prod;
1258 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1259 
1260 		i = rxq->ring.rsp_cons;
1261 		while ((i != rp)) {
1262 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1263 			memset(extras, 0, sizeof(rinfo.extras));
1264 
1265 			m = NULL;
1266 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1267 
1268 			if (__predict_false(err)) {
1269 				if (m)
1270 					(void )mbufq_enqueue(&mbufq_errq, m);
1271 				if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
1272 				continue;
1273 			}
1274 
1275 			m->m_pkthdr.rcvif = ifp;
1276 			if (rx->flags & NETRXF_data_validated) {
1277 				/*
1278 				 * According to mbuf(9) the correct way to tell
1279 				 * the stack that the checksum of an inbound
1280 				 * packet is correct, without it actually being
1281 				 * present (because the underlying interface
1282 				 * doesn't provide it), is to set the
1283 				 * CSUM_DATA_VALID and CSUM_PSEUDO_HDR flags,
1284 				 * and the csum_data field to 0xffff.
1285 				 */
1286 				m->m_pkthdr.csum_flags |= (CSUM_DATA_VALID
1287 				    | CSUM_PSEUDO_HDR);
1288 				m->m_pkthdr.csum_data = 0xffff;
1289 			}
1290 			if ((rx->flags & NETRXF_extra_info) != 0 &&
1291 			    (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1292 			    XEN_NETIF_EXTRA_TYPE_GSO)) {
1293 				m->m_pkthdr.tso_segsz =
1294 				extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1295 				m->m_pkthdr.csum_flags |= CSUM_TSO;
1296 			}
1297 
1298 			(void )mbufq_enqueue(&mbufq_rxq, m);
1299 		}
1300 
1301 		rxq->ring.rsp_cons = i;
1302 
1303 		xn_alloc_rx_buffers(rxq);
1304 
1305 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1306 	} while (work_to_do);
1307 
1308 	mbufq_drain(&mbufq_errq);
1309 	/*
1310 	 * Process all the mbufs after the remapping is complete.
1311 	 * Break the mbuf chain first though.
1312 	 */
1313 	while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1314 		if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1315 #if (defined(INET) || defined(INET6))
1316 		/* Use LRO if possible */
1317 		if ((if_getcapenable(ifp) & IFCAP_LRO) == 0 ||
1318 		    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1319 			/*
1320 			 * If LRO fails, pass up to the stack
1321 			 * directly.
1322 			 */
1323 			if_input(ifp, m);
1324 		}
1325 #else
1326 		if_input(ifp, m);
1327 #endif
1328 	}
1329 
1330 #if (defined(INET) || defined(INET6))
1331 	/*
1332 	 * Flush any outstanding LRO work
1333 	 */
1334 	tcp_lro_flush_all(lro);
1335 #endif
1336 }
1337 
1338 static void
1339 xn_txeof(struct netfront_txq *txq)
1340 {
1341 	RING_IDX i, prod;
1342 	unsigned short id;
1343 	if_t ifp;
1344 	netif_tx_response_t *txr;
1345 	struct mbuf *m;
1346 	struct netfront_info *np = txq->info;
1347 
1348 	XN_TX_LOCK_ASSERT(txq);
1349 
1350 	if (!netfront_carrier_ok(np))
1351 		return;
1352 
1353 	ifp = np->xn_ifp;
1354 
1355 	do {
1356 		prod = txq->ring.sring->rsp_prod;
1357 		rmb(); /* Ensure we see responses up to 'rp'. */
1358 
1359 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1360 			txr = RING_GET_RESPONSE(&txq->ring, i);
1361 			if (txr->status == NETIF_RSP_NULL)
1362 				continue;
1363 
1364 			if (txr->status != NETIF_RSP_OKAY) {
1365 				printf("%s: WARNING: response is %d!\n",
1366 				       __func__, txr->status);
1367 			}
1368 			id = txr->id;
1369 			m = txq->mbufs[id];
1370 			KASSERT(m != NULL, ("mbuf not found in chain"));
1371 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1372 				("mbuf already on the free list, but we're "
1373 				"trying to free it again!"));
1374 			M_ASSERTVALID(m);
1375 
1376 			if (__predict_false(gnttab_query_foreign_access(
1377 			    txq->grant_ref[id]) != 0)) {
1378 				panic("%s: grant id %u still in use by the "
1379 				    "backend", __func__, id);
1380 			}
1381 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1382 			gnttab_release_grant_reference(
1383 				&txq->gref_head, txq->grant_ref[id]);
1384 			txq->grant_ref[id] = GRANT_REF_INVALID;
1385 
1386 			txq->mbufs[id] = NULL;
1387 			add_id_to_freelist(txq->mbufs, id);
1388 			txq->mbufs_cnt--;
1389 			mbuf_release(m);
1390 			/* Only mark the txq active if we've freed up at least one slot to try */
1391 			if_setdrvflagbits(ifp, 0, IFF_DRV_OACTIVE);
1392 		}
1393 		txq->ring.rsp_cons = prod;
1394 
1395 		/*
1396 		 * Set a new event, then check for race with update of
1397 		 * tx_cons. Note that it is essential to schedule a
1398 		 * callback, no matter how few buffers are pending. Even if
1399 		 * there is space in the transmit ring, higher layers may
1400 		 * be blocked because too much data is outstanding: in such
1401 		 * cases notification from Xen is likely to be the only kick
1402 		 * that we'll get.
1403 		 */
1404 		txq->ring.sring->rsp_event =
1405 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1406 
1407 		mb();
1408 	} while (prod != txq->ring.sring->rsp_prod);
1409 
1410 	if (txq->full &&
1411 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1412 		txq->full = false;
1413 		xn_txq_start(txq);
1414 	}
1415 }
1416 
1417 static void
1418 xn_intr(void *xsc)
1419 {
1420 	struct netfront_txq *txq = xsc;
1421 	struct netfront_info *np = txq->info;
1422 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1423 
1424 	/* kick both tx and rx */
1425 	xn_rxq_intr(rxq);
1426 	xn_txq_intr(txq);
1427 }
1428 
1429 static void
1430 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1431     grant_ref_t ref)
1432 {
1433 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1434 
1435 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1436 	rxq->mbufs[new] = m;
1437 	rxq->grant_ref[new] = ref;
1438 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1439 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1440 	rxq->ring.req_prod_pvt++;
1441 }
1442 
1443 static int
1444 xn_get_extras(struct netfront_rxq *rxq,
1445     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1446 {
1447 	struct netif_extra_info *extra;
1448 
1449 	int err = 0;
1450 
1451 	do {
1452 		struct mbuf *m;
1453 		grant_ref_t ref;
1454 
1455 		if (__predict_false(*cons + 1 == rp)) {
1456 			err = EINVAL;
1457 			break;
1458 		}
1459 
1460 		extra = (struct netif_extra_info *)
1461 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1462 
1463 		if (__predict_false(!extra->type ||
1464 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1465 			err = EINVAL;
1466 		} else {
1467 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1468 		}
1469 
1470 		m = xn_get_rx_mbuf(rxq, *cons);
1471 		ref = xn_get_rx_ref(rxq,  *cons);
1472 		xn_move_rx_slot(rxq, m, ref);
1473 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1474 
1475 	return err;
1476 }
1477 
1478 static int
1479 xn_get_responses(struct netfront_rxq *rxq,
1480     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1481     struct mbuf  **list)
1482 {
1483 	struct netif_rx_response *rx = &rinfo->rx;
1484 	struct netif_extra_info *extras = rinfo->extras;
1485 	struct mbuf *m, *m0, *m_prev;
1486 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1487 	int frags = 1;
1488 	int err = 0;
1489 	u_long ret __diagused;
1490 
1491 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1492 
1493 	if (rx->flags & NETRXF_extra_info) {
1494 		err = xn_get_extras(rxq, extras, rp, cons);
1495 	}
1496 
1497 	if (m0 != NULL) {
1498 		m0->m_pkthdr.len = 0;
1499 		m0->m_next = NULL;
1500 	}
1501 
1502 	for (;;) {
1503 #if 0
1504 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1505 			rx->status, rx->offset, frags);
1506 #endif
1507 		if (__predict_false(rx->status < 0 ||
1508 			rx->offset + rx->status > PAGE_SIZE)) {
1509 			xn_move_rx_slot(rxq, m, ref);
1510 			if (m0 == m)
1511 				m0 = NULL;
1512 			m = NULL;
1513 			err = EINVAL;
1514 			goto next_skip_queue;
1515 		}
1516 
1517 		/*
1518 		 * This definitely indicates a bug, either in this driver or in
1519 		 * the backend driver. In future this should flag the bad
1520 		 * situation to the system controller to reboot the backed.
1521 		 */
1522 		if (ref == GRANT_REF_INVALID) {
1523 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1524 			err = EINVAL;
1525 			goto next;
1526 		}
1527 
1528 		ret = gnttab_end_foreign_access_ref(ref);
1529 		KASSERT(ret, ("Unable to end access to grant references"));
1530 
1531 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1532 
1533 next:
1534 		if (m == NULL)
1535 			break;
1536 
1537 		m->m_len = rx->status;
1538 		m->m_data += rx->offset;
1539 		m0->m_pkthdr.len += rx->status;
1540 
1541 next_skip_queue:
1542 		if (!(rx->flags & NETRXF_more_data))
1543 			break;
1544 
1545 		if (*cons + frags == rp) {
1546 			if (net_ratelimit())
1547 				WPRINTK("Need more frags\n");
1548 			err = ENOENT;
1549 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1550 			       __func__, *cons, frags, rp);
1551 			break;
1552 		}
1553 		/*
1554 		 * Note that m can be NULL, if rx->status < 0 or if
1555 		 * rx->offset + rx->status > PAGE_SIZE above.
1556 		 */
1557 		m_prev = m;
1558 
1559 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1560 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1561 
1562 		/*
1563 		 * m_prev == NULL can happen if rx->status < 0 or if
1564 		 * rx->offset + * rx->status > PAGE_SIZE above.
1565 		 */
1566 		if (m_prev != NULL)
1567 			m_prev->m_next = m;
1568 
1569 		/*
1570 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1571 		 * rx->status > PAGE_SIZE above.
1572 		 */
1573 		if (m0 == NULL)
1574 			m0 = m;
1575 		m->m_next = NULL;
1576 		ref = xn_get_rx_ref(rxq, *cons + frags);
1577 		frags++;
1578 	}
1579 	*list = m0;
1580 	*cons += frags;
1581 
1582 	return (err);
1583 }
1584 
1585 /**
1586  * Given an mbuf chain, make sure we have enough room and then push
1587  * it onto the transmit ring.
1588  */
1589 static int
1590 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1591 {
1592 	struct netfront_info *np = txq->info;
1593 	if_t ifp = np->xn_ifp;
1594 	int otherend_id, error, nfrags;
1595 	bus_dma_segment_t *segs = txq->segs;
1596 	struct mbuf_xennet *tag;
1597 	bus_dmamap_t map;
1598 	unsigned int i;
1599 
1600 	KASSERT(!SLIST_EMPTY(&txq->tags), ("no tags available"));
1601 	tag = SLIST_FIRST(&txq->tags);
1602 	SLIST_REMOVE_HEAD(&txq->tags, next);
1603 	KASSERT(tag->count == 0, ("tag already in-use"));
1604 	map = tag->dma_map;
1605 	error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1606 	    &nfrags, 0);
1607 	if (error == EFBIG || nfrags > np->maxfrags) {
1608 		struct mbuf *m;
1609 
1610 		bus_dmamap_unload(np->dma_tag, map);
1611 		m = m_defrag(m_head, M_NOWAIT);
1612 		if (!m) {
1613 			/*
1614 			 * Defrag failed, so free the mbuf and
1615 			 * therefore drop the packet.
1616 			 */
1617 			SLIST_INSERT_HEAD(&txq->tags, tag, next);
1618 			m_freem(m_head);
1619 			return (EMSGSIZE);
1620 		}
1621 		m_head = m;
1622 		error = bus_dmamap_load_mbuf_sg(np->dma_tag, map, m_head, segs,
1623 		    &nfrags, 0);
1624 		if (error != 0 || nfrags > np->maxfrags) {
1625 			bus_dmamap_unload(np->dma_tag, map);
1626 			SLIST_INSERT_HEAD(&txq->tags, tag, next);
1627 			m_freem(m_head);
1628 			return (error ?: EFBIG);
1629 		}
1630 	} else if (error != 0) {
1631 		SLIST_INSERT_HEAD(&txq->tags, tag, next);
1632 		m_freem(m_head);
1633 		return (error);
1634 	}
1635 
1636 	/**
1637 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1638 	 * of mbufs longer than Linux can handle.  Make sure we don't
1639 	 * pass a too-long chain over to the other side by dropping the
1640 	 * packet.  It doesn't look like there is currently a way to
1641 	 * tell the TCP stack to generate a shorter chain of packets.
1642 	 */
1643 	if (nfrags > MAX_TX_REQ_FRAGS) {
1644 #ifdef DEBUG
1645 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1646 		       "won't be able to handle it, dropping\n",
1647 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1648 #endif
1649 		SLIST_INSERT_HEAD(&txq->tags, tag, next);
1650 		bus_dmamap_unload(np->dma_tag, map);
1651 		m_freem(m_head);
1652 		return (EMSGSIZE);
1653 	}
1654 
1655 	/*
1656 	 * This check should be redundant.  We've already verified that we
1657 	 * have enough slots in the ring to handle a packet of maximum
1658 	 * size, and that our packet is less than the maximum size.  Keep
1659 	 * it in here as an assert for now just to make certain that
1660 	 * chain_cnt is accurate.
1661 	 */
1662 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1663 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1664 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1665                     (int) nfrags, (int) NET_TX_RING_SIZE));
1666 
1667 	/*
1668 	 * Start packing the mbufs in this chain into
1669 	 * the fragment pointers. Stop when we run out
1670 	 * of fragments or hit the end of the mbuf chain.
1671 	 */
1672 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1673 	m_tag_prepend(m_head, &tag->tag);
1674 	for (i = 0; i < nfrags; i++) {
1675 		netif_tx_request_t *tx;
1676 		uintptr_t id;
1677 		grant_ref_t ref;
1678 		u_long mfn; /* XXX Wrong type? */
1679 
1680 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1681 		id = get_id_from_freelist(txq->mbufs);
1682 		if (id == 0)
1683 			panic("%s: was allocated the freelist head!\n",
1684 			    __func__);
1685 		txq->mbufs_cnt++;
1686 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1687 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1688 			    __func__);
1689 		mbuf_grab(m_head);
1690 		txq->mbufs[id] = m_head;
1691 		tx->id = id;
1692 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1693 		KASSERT((short)ref >= 0, ("Negative ref"));
1694 		mfn = atop(segs[i].ds_addr);
1695 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1696 		    mfn, GNTMAP_readonly);
1697 		tx->gref = txq->grant_ref[id] = ref;
1698 		tx->offset = segs[i].ds_addr & PAGE_MASK;
1699 		KASSERT(tx->offset + segs[i].ds_len <= PAGE_SIZE,
1700 		    ("mbuf segment crosses a page boundary"));
1701 		tx->flags = 0;
1702 		if (i == 0) {
1703 			/*
1704 			 * The first fragment has the entire packet
1705 			 * size, subsequent fragments have just the
1706 			 * fragment size. The backend works out the
1707 			 * true size of the first fragment by
1708 			 * subtracting the sizes of the other
1709 			 * fragments.
1710 			 */
1711 			tx->size = m_head->m_pkthdr.len;
1712 
1713 			/*
1714 			 * The first fragment contains the checksum flags
1715 			 * and is optionally followed by extra data for
1716 			 * TSO etc.
1717 			 */
1718 			/**
1719 			 * CSUM_TSO requires checksum offloading.
1720 			 * Some versions of FreeBSD fail to
1721 			 * set CSUM_TCP in the CSUM_TSO case,
1722 			 * so we have to test for CSUM_TSO
1723 			 * explicitly.
1724 			 */
1725 			if (m_head->m_pkthdr.csum_flags
1726 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1727 				tx->flags |= (NETTXF_csum_blank
1728 				    | NETTXF_data_validated);
1729 			}
1730 			if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
1731 				struct netif_extra_info *gso =
1732 					(struct netif_extra_info *)
1733 					RING_GET_REQUEST(&txq->ring,
1734 							 ++txq->ring.req_prod_pvt);
1735 
1736 				tx->flags |= NETTXF_extra_info;
1737 
1738 				gso->u.gso.size = m_head->m_pkthdr.tso_segsz;
1739 				gso->u.gso.type =
1740 					XEN_NETIF_GSO_TYPE_TCPV4;
1741 				gso->u.gso.pad = 0;
1742 				gso->u.gso.features = 0;
1743 
1744 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1745 				gso->flags = 0;
1746 			}
1747 		} else {
1748 			tx->size = segs[i].ds_len;
1749 		}
1750 		if (i != nfrags - 1)
1751 			tx->flags |= NETTXF_more_data;
1752 
1753 		txq->ring.req_prod_pvt++;
1754 	}
1755 	bus_dmamap_sync(np->dma_tag, map, BUS_DMASYNC_PREWRITE);
1756 	BPF_MTAP(ifp, m_head);
1757 
1758 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1759 	if_inc_counter(ifp, IFCOUNTER_OBYTES, m_head->m_pkthdr.len);
1760 	if (m_head->m_flags & M_MCAST)
1761 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
1762 
1763 	xn_txeof(txq);
1764 
1765 	return (0);
1766 }
1767 
1768 /* equivalent of network_open() in Linux */
1769 static void
1770 xn_ifinit_locked(struct netfront_info *np)
1771 {
1772 	if_t ifp;
1773 	int i;
1774 	struct netfront_rxq *rxq;
1775 
1776 	XN_LOCK_ASSERT(np);
1777 
1778 	ifp = np->xn_ifp;
1779 
1780 	if (if_getdrvflags(ifp) & IFF_DRV_RUNNING || !netfront_carrier_ok(np))
1781 		return;
1782 
1783 	xn_stop(np);
1784 
1785 	for (i = 0; i < np->num_queues; i++) {
1786 		rxq = &np->rxq[i];
1787 		XN_RX_LOCK(rxq);
1788 		xn_alloc_rx_buffers(rxq);
1789 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1790 		if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1791 			xn_rxeof(rxq);
1792 		XN_RX_UNLOCK(rxq);
1793 	}
1794 
1795 	if_setdrvflagbits(ifp, IFF_DRV_RUNNING, 0);
1796 	if_setdrvflagbits(ifp, 0, IFF_DRV_OACTIVE);
1797 	if_link_state_change(ifp, LINK_STATE_UP);
1798 }
1799 
1800 static void
1801 xn_ifinit(void *xsc)
1802 {
1803 	struct netfront_info *sc = xsc;
1804 
1805 	XN_LOCK(sc);
1806 	xn_ifinit_locked(sc);
1807 	XN_UNLOCK(sc);
1808 }
1809 
1810 static int
1811 xn_ioctl(if_t ifp, u_long cmd, caddr_t data)
1812 {
1813 	struct netfront_info *sc = if_getsoftc(ifp);
1814 	struct ifreq *ifr = (struct ifreq *) data;
1815 	device_t dev;
1816 #ifdef INET
1817 	struct ifaddr *ifa = (struct ifaddr *)data;
1818 #endif
1819 	int mask, error = 0, reinit;
1820 
1821 	dev = sc->xbdev;
1822 
1823 	switch(cmd) {
1824 	case SIOCSIFADDR:
1825 #ifdef INET
1826 		XN_LOCK(sc);
1827 		if (ifa->ifa_addr->sa_family == AF_INET) {
1828 			if_setflagbits(ifp, IFF_UP, 0);
1829 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
1830 				xn_ifinit_locked(sc);
1831 			arp_ifinit(ifp, ifa);
1832 			XN_UNLOCK(sc);
1833 		} else {
1834 			XN_UNLOCK(sc);
1835 #endif
1836 			error = ether_ioctl(ifp, cmd, data);
1837 #ifdef INET
1838 		}
1839 #endif
1840 		break;
1841 	case SIOCSIFMTU:
1842 		if (if_getmtu(ifp) == ifr->ifr_mtu)
1843 			break;
1844 
1845 		if_setmtu(ifp, ifr->ifr_mtu);
1846 		if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING);
1847 		xn_ifinit(sc);
1848 		break;
1849 	case SIOCSIFFLAGS:
1850 		XN_LOCK(sc);
1851 		if (if_getflags(ifp) & IFF_UP) {
1852 			/*
1853 			 * If only the state of the PROMISC flag changed,
1854 			 * then just use the 'set promisc mode' command
1855 			 * instead of reinitializing the entire NIC. Doing
1856 			 * a full re-init means reloading the firmware and
1857 			 * waiting for it to start up, which may take a
1858 			 * second or two.
1859 			 */
1860 			xn_ifinit_locked(sc);
1861 		} else {
1862 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
1863 				xn_stop(sc);
1864 			}
1865 		}
1866 		sc->xn_if_flags = if_getflags(ifp);
1867 		XN_UNLOCK(sc);
1868 		break;
1869 	case SIOCSIFCAP:
1870 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
1871 		reinit = 0;
1872 
1873 		if (mask & IFCAP_TXCSUM) {
1874 			if_togglecapenable(ifp, IFCAP_TXCSUM);
1875 			if_togglehwassist(ifp, XN_CSUM_FEATURES);
1876 		}
1877 		if (mask & IFCAP_TSO4) {
1878 			if_togglecapenable(ifp, IFCAP_TSO4);
1879 			if_togglehwassist(ifp, CSUM_TSO);
1880 		}
1881 
1882 		if (mask & (IFCAP_RXCSUM | IFCAP_LRO)) {
1883 			/* These Rx features require us to renegotiate. */
1884 			reinit = 1;
1885 
1886 			if (mask & IFCAP_RXCSUM)
1887 				if_togglecapenable(ifp, IFCAP_RXCSUM);
1888 			if (mask & IFCAP_LRO)
1889 				if_togglecapenable(ifp, IFCAP_LRO);
1890 		}
1891 
1892 		if (reinit == 0)
1893 			break;
1894 
1895 		/*
1896 		 * We must reset the interface so the backend picks up the
1897 		 * new features.
1898 		 */
1899 		device_printf(sc->xbdev,
1900 		    "performing interface reset due to feature change\n");
1901 		XN_LOCK(sc);
1902 		netfront_carrier_off(sc);
1903 		sc->xn_reset = true;
1904 		/*
1905 		 * NB: the pending packet queue is not flushed, since
1906 		 * the interface should still support the old options.
1907 		 */
1908 		XN_UNLOCK(sc);
1909 		/*
1910 		 * Delete the xenstore nodes that export features.
1911 		 *
1912 		 * NB: There's a xenbus state called
1913 		 * "XenbusStateReconfiguring", which is what we should set
1914 		 * here. Sadly none of the backends know how to handle it,
1915 		 * and simply disconnect from the frontend, so we will just
1916 		 * switch back to XenbusStateInitialising in order to force
1917 		 * a reconnection.
1918 		 */
1919 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-gso-tcpv4");
1920 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-no-csum-offload");
1921 		xenbus_set_state(dev, XenbusStateClosing);
1922 
1923 		/*
1924 		 * Wait for the frontend to reconnect before returning
1925 		 * from the ioctl. 30s should be more than enough for any
1926 		 * sane backend to reconnect.
1927 		 */
1928 		error = tsleep(sc, 0, "xn_rst", 30*hz);
1929 		break;
1930 	case SIOCADDMULTI:
1931 	case SIOCDELMULTI:
1932 		break;
1933 	case SIOCSIFMEDIA:
1934 	case SIOCGIFMEDIA:
1935 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1936 		break;
1937 	default:
1938 		error = ether_ioctl(ifp, cmd, data);
1939 	}
1940 
1941 	return (error);
1942 }
1943 
1944 static void
1945 xn_stop(struct netfront_info *sc)
1946 {
1947 	if_t ifp;
1948 
1949 	XN_LOCK_ASSERT(sc);
1950 
1951 	ifp = sc->xn_ifp;
1952 
1953 	if_setdrvflagbits(ifp, 0, IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1954 	if_link_state_change(ifp, LINK_STATE_DOWN);
1955 }
1956 
1957 static void
1958 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1959 {
1960 	int requeue_idx, i;
1961 	grant_ref_t ref;
1962 	netif_rx_request_t *req;
1963 
1964 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1965 		struct mbuf *m;
1966 		u_long pfn;
1967 
1968 		if (rxq->mbufs[i] == NULL)
1969 			continue;
1970 
1971 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1972 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1973 
1974 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1975 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1976 
1977 		gnttab_grant_foreign_access_ref(ref,
1978 		    xenbus_get_otherend_id(rxq->info->xbdev),
1979 		    pfn, 0);
1980 
1981 		req->gref = ref;
1982 		req->id   = requeue_idx;
1983 
1984 		requeue_idx++;
1985 	}
1986 
1987 	rxq->ring.req_prod_pvt = requeue_idx;
1988 }
1989 
1990 /* START of Xenolinux helper functions adapted to FreeBSD */
1991 static int
1992 xn_connect(struct netfront_info *np)
1993 {
1994 	int i, error;
1995 	u_int feature_rx_copy;
1996 	struct netfront_rxq *rxq;
1997 	struct netfront_txq *txq;
1998 
1999 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2000 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
2001 	if (error != 0)
2002 		feature_rx_copy = 0;
2003 
2004 	/* We only support rx copy. */
2005 	if (!feature_rx_copy)
2006 		return (EPROTONOSUPPORT);
2007 
2008 	/* Recovery procedure: */
2009 	error = talk_to_backend(np->xbdev, np);
2010 	if (error != 0)
2011 		return (error);
2012 
2013 	/* Step 1: Reinitialise variables. */
2014 	xn_query_features(np);
2015 	xn_configure_features(np);
2016 
2017 	/* Step 2: Release TX buffer */
2018 	for (i = 0; i < np->num_queues; i++) {
2019 		txq = &np->txq[i];
2020 		xn_release_tx_bufs(txq);
2021 	}
2022 
2023 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
2024 	for (i = 0; i < np->num_queues; i++) {
2025 		rxq = &np->rxq[i];
2026 		xn_rebuild_rx_bufs(rxq);
2027 	}
2028 
2029 	/* Step 4: All public and private state should now be sane.  Get
2030 	 * ready to start sending and receiving packets and give the driver
2031 	 * domain a kick because we've probably just requeued some
2032 	 * packets.
2033 	 */
2034 	netfront_carrier_on(np);
2035 	wakeup(np);
2036 
2037 	return (0);
2038 }
2039 
2040 static void
2041 xn_kick_rings(struct netfront_info *np)
2042 {
2043 	struct netfront_rxq *rxq;
2044 	struct netfront_txq *txq;
2045 	int i;
2046 
2047 	for (i = 0; i < np->num_queues; i++) {
2048 		txq = &np->txq[i];
2049 		rxq = &np->rxq[i];
2050 		xen_intr_signal(txq->xen_intr_handle);
2051 		XN_TX_LOCK(txq);
2052 		xn_txeof(txq);
2053 		XN_TX_UNLOCK(txq);
2054 		XN_RX_LOCK(rxq);
2055 		xn_alloc_rx_buffers(rxq);
2056 		XN_RX_UNLOCK(rxq);
2057 	}
2058 }
2059 
2060 static void
2061 xn_query_features(struct netfront_info *np)
2062 {
2063 	int val;
2064 
2065 	device_printf(np->xbdev, "backend features:");
2066 
2067 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2068 		"feature-sg", NULL, "%d", &val) != 0)
2069 		val = 0;
2070 
2071 	np->maxfrags = 1;
2072 	if (val) {
2073 		np->maxfrags = MAX_TX_REQ_FRAGS;
2074 		printf(" feature-sg");
2075 	}
2076 
2077 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2078 		"feature-gso-tcpv4", NULL, "%d", &val) != 0)
2079 		val = 0;
2080 
2081 	if_setcapabilitiesbit(np->xn_ifp, 0, IFCAP_TSO4 | IFCAP_LRO);
2082 	if (val) {
2083 		if_setcapabilitiesbit(np->xn_ifp, IFCAP_TSO4 | IFCAP_LRO, 0);
2084 		printf(" feature-gso-tcp4");
2085 	}
2086 
2087 	/*
2088 	 * HW CSUM offload is assumed to be available unless
2089 	 * feature-no-csum-offload is set in xenstore.
2090 	 */
2091 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2092 		"feature-no-csum-offload", NULL, "%d", &val) != 0)
2093 		val = 0;
2094 
2095 	if_setcapabilitiesbit(np->xn_ifp, IFCAP_HWCSUM, 0);
2096 	if (val) {
2097 		if_setcapabilitiesbit(np->xn_ifp, 0, IFCAP_HWCSUM);
2098 		printf(" feature-no-csum-offload");
2099 	}
2100 
2101 	printf("\n");
2102 }
2103 
2104 static int
2105 xn_configure_features(struct netfront_info *np)
2106 {
2107 	int err, cap_enabled;
2108 #if (defined(INET) || defined(INET6))
2109 	int i;
2110 #endif
2111 	if_t ifp;
2112 
2113 	ifp = np->xn_ifp;
2114 	err = 0;
2115 
2116 	if ((if_getcapenable(ifp) & if_getcapabilities(ifp)) == if_getcapenable(ifp)) {
2117 		/* Current options are available, no need to do anything. */
2118 		return (0);
2119 	}
2120 
2121 	/* Try to preserve as many options as possible. */
2122 	cap_enabled = if_getcapenable(ifp);
2123 	if_setcapenable(ifp, 0);
2124 	if_sethwassist(ifp, 0);
2125 
2126 #if (defined(INET) || defined(INET6))
2127 	if ((cap_enabled & IFCAP_LRO) != 0)
2128 		for (i = 0; i < np->num_queues; i++)
2129 			tcp_lro_free(&np->rxq[i].lro);
2130 	if (xn_enable_lro &&
2131 	    (if_getcapabilities(ifp) & cap_enabled & IFCAP_LRO) != 0) {
2132 	    	if_setcapenablebit(ifp, IFCAP_LRO, 0);
2133 		for (i = 0; i < np->num_queues; i++) {
2134 			err = tcp_lro_init(&np->rxq[i].lro);
2135 			if (err != 0) {
2136 				device_printf(np->xbdev,
2137 				    "LRO initialization failed\n");
2138 				if_setcapenablebit(ifp, 0, IFCAP_LRO);
2139 				break;
2140 			}
2141 			np->rxq[i].lro.ifp = ifp;
2142 		}
2143 	}
2144 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_TSO4) != 0) {
2145 		if_setcapenablebit(ifp, IFCAP_TSO4, 0);
2146 		if_sethwassistbits(ifp, CSUM_TSO, 0);
2147 	}
2148 #endif
2149 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_TXCSUM) != 0) {
2150 		if_setcapenablebit(ifp, IFCAP_TXCSUM, 0);
2151 		if_sethwassistbits(ifp, XN_CSUM_FEATURES, 0);
2152 	}
2153 	if ((if_getcapabilities(ifp) & cap_enabled & IFCAP_RXCSUM) != 0)
2154 		if_setcapenablebit(ifp, IFCAP_RXCSUM, 0);
2155 
2156 	return (err);
2157 }
2158 
2159 static int
2160 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2161 {
2162 	struct netfront_info *np;
2163 	if_t ifp;
2164 	struct buf_ring *br;
2165 	int error, notify;
2166 
2167 	np = txq->info;
2168 	br = txq->br;
2169 	ifp = np->xn_ifp;
2170 	error = 0;
2171 
2172 	XN_TX_LOCK_ASSERT(txq);
2173 
2174 	if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0 ||
2175 	    !netfront_carrier_ok(np)) {
2176 		if (m != NULL)
2177 			error = drbr_enqueue(ifp, br, m);
2178 		return (error);
2179 	}
2180 
2181 	if (m != NULL) {
2182 		error = drbr_enqueue(ifp, br, m);
2183 		if (error != 0)
2184 			return (error);
2185 	}
2186 
2187 	while ((m = drbr_peek(ifp, br)) != NULL) {
2188 		if (!xn_tx_slot_available(txq)) {
2189 			drbr_putback(ifp, br, m);
2190 			break;
2191 		}
2192 
2193 		error = xn_assemble_tx_request(txq, m);
2194 		/* xn_assemble_tx_request always consumes the mbuf*/
2195 		if (error != 0) {
2196 			drbr_advance(ifp, br);
2197 			break;
2198 		}
2199 
2200 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2201 		if (notify)
2202 			xen_intr_signal(txq->xen_intr_handle);
2203 
2204 		drbr_advance(ifp, br);
2205 	}
2206 
2207 	if (RING_FULL(&txq->ring))
2208 		txq->full = true;
2209 
2210 	return (0);
2211 }
2212 
2213 static int
2214 xn_txq_mq_start(if_t ifp, struct mbuf *m)
2215 {
2216 	struct netfront_info *np;
2217 	struct netfront_txq *txq;
2218 	int i, npairs, error;
2219 
2220 	np = if_getsoftc(ifp);
2221 	npairs = np->num_queues;
2222 
2223 	if (!netfront_carrier_ok(np))
2224 		return (ENOBUFS);
2225 
2226 	KASSERT(npairs != 0, ("called with 0 available queues"));
2227 
2228 	/* check if flowid is set */
2229 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2230 		i = m->m_pkthdr.flowid % npairs;
2231 	else
2232 		i = curcpu % npairs;
2233 
2234 	txq = &np->txq[i];
2235 
2236 	if (XN_TX_TRYLOCK(txq) != 0) {
2237 		error = xn_txq_mq_start_locked(txq, m);
2238 		XN_TX_UNLOCK(txq);
2239 	} else {
2240 		error = drbr_enqueue(ifp, txq->br, m);
2241 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2242 	}
2243 
2244 	return (error);
2245 }
2246 
2247 static void
2248 xn_qflush(if_t ifp)
2249 {
2250 	struct netfront_info *np;
2251 	struct netfront_txq *txq;
2252 	struct mbuf *m;
2253 	int i;
2254 
2255 	np = if_getsoftc(ifp);
2256 
2257 	for (i = 0; i < np->num_queues; i++) {
2258 		txq = &np->txq[i];
2259 
2260 		XN_TX_LOCK(txq);
2261 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2262 			m_freem(m);
2263 		XN_TX_UNLOCK(txq);
2264 	}
2265 
2266 	if_qflush(ifp);
2267 }
2268 
2269 /**
2270  * Create a network device.
2271  * @param dev  Newbus device representing this virtual NIC.
2272  */
2273 int
2274 create_netdev(device_t dev)
2275 {
2276 	struct netfront_info *np;
2277 	int err;
2278 	if_t ifp;
2279 
2280 	np = device_get_softc(dev);
2281 
2282 	np->xbdev         = dev;
2283 
2284 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2285 
2286 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2287 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2288 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2289 
2290 	err = xen_net_read_mac(dev, np->mac);
2291 	if (err != 0)
2292 		goto error;
2293 
2294 	/* Set up ifnet structure */
2295 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2296 	if_setsoftc(ifp, np);
2297 	if_initname(ifp, "xn",  device_get_unit(dev));
2298 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
2299 	if_setioctlfn(ifp, xn_ioctl);
2300 
2301 	if_settransmitfn(ifp, xn_txq_mq_start);
2302 	if_setqflushfn(ifp, xn_qflush);
2303 
2304 	if_setinitfn(ifp, xn_ifinit);
2305 
2306 	if_sethwassist(ifp, XN_CSUM_FEATURES);
2307 	/* Enable all supported features at device creation. */
2308 	if_setcapabilities(ifp, IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_LRO);
2309 	if_setcapenable(ifp, if_getcapabilities(ifp));
2310 
2311 	if_sethwtsomax(ifp, 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN));
2312 	if_sethwtsomaxsegcount(ifp, MAX_TX_REQ_FRAGS);
2313 	if_sethwtsomaxsegsize(ifp, PAGE_SIZE);
2314 
2315 	ether_ifattach(ifp, np->mac);
2316 	netfront_carrier_off(np);
2317 
2318 	err = bus_dma_tag_create(
2319 	    bus_get_dma_tag(dev),		/* parent */
2320 	    1, PAGE_SIZE,			/* algnmnt, boundary */
2321 	    BUS_SPACE_MAXADDR,			/* lowaddr */
2322 	    BUS_SPACE_MAXADDR,			/* highaddr */
2323 	    NULL, NULL,				/* filter, filterarg */
2324 	    PAGE_SIZE * MAX_TX_REQ_FRAGS,	/* max request size */
2325 	    MAX_TX_REQ_FRAGS,			/* max segments */
2326 	    PAGE_SIZE,				/* maxsegsize */
2327 	    BUS_DMA_ALLOCNOW,			/* flags */
2328 	    NULL, NULL,				/* lockfunc, lockarg */
2329 	    &np->dma_tag);
2330 
2331 	return (err);
2332 
2333 error:
2334 	KASSERT(err != 0, ("Error path with no error code specified"));
2335 	return (err);
2336 }
2337 
2338 static int
2339 netfront_detach(device_t dev)
2340 {
2341 	struct netfront_info *info = device_get_softc(dev);
2342 
2343 	DPRINTK("%s\n", xenbus_get_node(dev));
2344 
2345 	netif_free(info);
2346 
2347 	return 0;
2348 }
2349 
2350 static void
2351 netif_free(struct netfront_info *np)
2352 {
2353 
2354 	XN_LOCK(np);
2355 	xn_stop(np);
2356 	XN_UNLOCK(np);
2357 	netif_disconnect_backend(np);
2358 	ether_ifdetach(np->xn_ifp);
2359 	free(np->rxq, M_DEVBUF);
2360 	free(np->txq, M_DEVBUF);
2361 	if_free(np->xn_ifp);
2362 	np->xn_ifp = NULL;
2363 	ifmedia_removeall(&np->sc_media);
2364 	bus_dma_tag_destroy(np->dma_tag);
2365 }
2366 
2367 static void
2368 netif_disconnect_backend(struct netfront_info *np)
2369 {
2370 	u_int i;
2371 
2372 	for (i = 0; i < np->num_queues; i++) {
2373 		XN_RX_LOCK(&np->rxq[i]);
2374 		XN_TX_LOCK(&np->txq[i]);
2375 	}
2376 	netfront_carrier_off(np);
2377 	for (i = 0; i < np->num_queues; i++) {
2378 		XN_RX_UNLOCK(&np->rxq[i]);
2379 		XN_TX_UNLOCK(&np->txq[i]);
2380 	}
2381 
2382 	for (i = 0; i < np->num_queues; i++) {
2383 		disconnect_rxq(&np->rxq[i]);
2384 		disconnect_txq(&np->txq[i]);
2385 	}
2386 }
2387 
2388 static int
2389 xn_ifmedia_upd(if_t ifp)
2390 {
2391 
2392 	return (0);
2393 }
2394 
2395 static void
2396 xn_ifmedia_sts(if_t ifp, struct ifmediareq *ifmr)
2397 {
2398 
2399 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2400 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2401 }
2402 
2403 /* ** Driver registration ** */
2404 static device_method_t netfront_methods[] = {
2405 	/* Device interface */
2406 	DEVMETHOD(device_probe,         netfront_probe),
2407 	DEVMETHOD(device_attach,        netfront_attach),
2408 	DEVMETHOD(device_detach,        netfront_detach),
2409 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2410 	DEVMETHOD(device_suspend,       netfront_suspend),
2411 	DEVMETHOD(device_resume,        netfront_resume),
2412 
2413 	/* Xenbus interface */
2414 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2415 
2416 	DEVMETHOD_END
2417 };
2418 
2419 static driver_t netfront_driver = {
2420 	"xn",
2421 	netfront_methods,
2422 	sizeof(struct netfront_info),
2423 };
2424 
2425 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, NULL, NULL);
2426