xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision d039b0700bf099d48acf5f07a933a0c17f738c56)
1 /*-
2  * Copyright (c) 2004-2006 Kip Macy
3  * Copyright (c) 2015 Wei Liu <wei.liu2@citrix.com>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
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 <xen/interface/memory.h>
69 #include <xen/interface/io/netif.h>
70 #include <xen/xenbus/xenbusvar.h>
71 
72 #include "xenbus_if.h"
73 
74 /* Features supported by all backends.  TSO and LRO can be negotiated */
75 #define XN_CSUM_FEATURES	(CSUM_TCP | CSUM_UDP)
76 
77 #define NET_TX_RING_SIZE __RING_SIZE((netif_tx_sring_t *)0, PAGE_SIZE)
78 #define NET_RX_RING_SIZE __RING_SIZE((netif_rx_sring_t *)0, PAGE_SIZE)
79 
80 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
81 
82 /*
83  * Should the driver do LRO on the RX end
84  *  this can be toggled on the fly, but the
85  *  interface must be reset (down/up) for it
86  *  to take effect.
87  */
88 static int xn_enable_lro = 1;
89 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
90 
91 /*
92  * Number of pairs of queues.
93  */
94 static unsigned long xn_num_queues = 4;
95 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
96 
97 /**
98  * \brief The maximum allowed data fragments in a single transmit
99  *        request.
100  *
101  * This limit is imposed by the backend driver.  We assume here that
102  * we are dealing with a Linux driver domain and have set our limit
103  * to mirror the Linux MAX_SKB_FRAGS constant.
104  */
105 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
106 
107 #define RX_COPY_THRESHOLD 256
108 
109 #define net_ratelimit() 0
110 
111 struct netfront_rxq;
112 struct netfront_txq;
113 struct netfront_info;
114 struct netfront_rx_info;
115 
116 static void xn_txeof(struct netfront_txq *);
117 static void xn_rxeof(struct netfront_rxq *);
118 static void xn_alloc_rx_buffers(struct netfront_rxq *);
119 static void xn_alloc_rx_buffers_callout(void *arg);
120 
121 static void xn_release_rx_bufs(struct netfront_rxq *);
122 static void xn_release_tx_bufs(struct netfront_txq *);
123 
124 static void xn_rxq_intr(struct netfront_rxq *);
125 static void xn_txq_intr(struct netfront_txq *);
126 static void xn_intr(void *);
127 static inline int xn_count_frags(struct mbuf *m);
128 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
129 static int xn_ioctl(struct ifnet *, u_long, caddr_t);
130 static void xn_ifinit_locked(struct netfront_info *);
131 static void xn_ifinit(void *);
132 static void xn_stop(struct netfront_info *);
133 static void xn_query_features(struct netfront_info *np);
134 static int xn_configure_features(struct netfront_info *np);
135 static void netif_free(struct netfront_info *info);
136 static int netfront_detach(device_t dev);
137 
138 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
139 static int xn_txq_mq_start(struct ifnet *, struct mbuf *);
140 
141 static int talk_to_backend(device_t dev, struct netfront_info *info);
142 static int create_netdev(device_t dev);
143 static void netif_disconnect_backend(struct netfront_info *info);
144 static int setup_device(device_t dev, struct netfront_info *info,
145     unsigned long);
146 static int xn_ifmedia_upd(struct ifnet *ifp);
147 static void xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr);
148 
149 int xn_connect(struct netfront_info *);
150 
151 static int xn_get_responses(struct netfront_rxq *,
152     struct netfront_rx_info *, RING_IDX, RING_IDX *,
153     struct mbuf **);
154 
155 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
156 
157 #define INVALID_P2M_ENTRY (~0UL)
158 
159 struct xn_rx_stats
160 {
161 	u_long	rx_packets;	/* total packets received	*/
162 	u_long	rx_bytes;	/* total bytes received 	*/
163 	u_long	rx_errors;	/* bad packets received		*/
164 };
165 
166 struct xn_tx_stats
167 {
168 	u_long	tx_packets;	/* total packets transmitted	*/
169 	u_long	tx_bytes;	/* total bytes transmitted	*/
170 	u_long	tx_errors;	/* packet transmit problems	*/
171 };
172 
173 #define XN_QUEUE_NAME_LEN  8	/* xn{t,r}x_%u, allow for two digits */
174 struct netfront_rxq {
175 	struct netfront_info 	*info;
176 	u_int			id;
177 	char			name[XN_QUEUE_NAME_LEN];
178 	struct mtx		lock;
179 
180 	int			ring_ref;
181 	netif_rx_front_ring_t 	ring;
182 	xen_intr_handle_t	xen_intr_handle;
183 
184 	grant_ref_t 		gref_head;
185 	grant_ref_t 		grant_ref[NET_TX_RING_SIZE + 1];
186 
187 	struct mbuf		*mbufs[NET_RX_RING_SIZE + 1];
188 
189 	struct lro_ctrl		lro;
190 
191 	struct callout		rx_refill;
192 
193 	struct xn_rx_stats	stats;
194 };
195 
196 struct netfront_txq {
197 	struct netfront_info 	*info;
198 	u_int 			id;
199 	char			name[XN_QUEUE_NAME_LEN];
200 	struct mtx		lock;
201 
202 	int			ring_ref;
203 	netif_tx_front_ring_t	ring;
204 	xen_intr_handle_t 	xen_intr_handle;
205 
206 	grant_ref_t		gref_head;
207 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
208 
209 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
210 	int			mbufs_cnt;
211 	struct buf_ring		*br;
212 
213 	struct taskqueue 	*tq;
214 	struct task       	defrtask;
215 
216 	bool			full;
217 
218 	struct xn_tx_stats	stats;
219 };
220 
221 struct netfront_info {
222 	struct ifnet 		*xn_ifp;
223 
224 	struct mtx   		sc_lock;
225 
226 	u_int  num_queues;
227 	struct netfront_rxq 	*rxq;
228 	struct netfront_txq 	*txq;
229 
230 	u_int			carrier;
231 	u_int			maxfrags;
232 
233 	device_t		xbdev;
234 	uint8_t			mac[ETHER_ADDR_LEN];
235 
236 	int			xn_if_flags;
237 
238 	struct ifmedia		sc_media;
239 
240 	bool			xn_resume;
241 };
242 
243 struct netfront_rx_info {
244 	struct netif_rx_response rx;
245 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
246 };
247 
248 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
249 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
250 
251 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
252 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
253 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
254 
255 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
256 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
257 
258 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
259 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
260 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
261 
262 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
263 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
264 #define netfront_carrier_ok(netif)	((netif)->carrier)
265 
266 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
267 
268 static inline void
269 add_id_to_freelist(struct mbuf **list, uintptr_t id)
270 {
271 
272 	KASSERT(id != 0,
273 		("%s: the head item (0) must always be free.", __func__));
274 	list[id] = list[0];
275 	list[0]  = (struct mbuf *)id;
276 }
277 
278 static inline unsigned short
279 get_id_from_freelist(struct mbuf **list)
280 {
281 	uintptr_t id;
282 
283 	id = (uintptr_t)list[0];
284 	KASSERT(id != 0,
285 		("%s: the head item (0) must always remain free.", __func__));
286 	list[0] = list[id];
287 	return (id);
288 }
289 
290 static inline int
291 xn_rxidx(RING_IDX idx)
292 {
293 
294 	return idx & (NET_RX_RING_SIZE - 1);
295 }
296 
297 static inline struct mbuf *
298 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
299 {
300 	int i;
301 	struct mbuf *m;
302 
303 	i = xn_rxidx(ri);
304 	m = rxq->mbufs[i];
305 	rxq->mbufs[i] = NULL;
306 	return (m);
307 }
308 
309 static inline grant_ref_t
310 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
311 {
312 	int i = xn_rxidx(ri);
313 	grant_ref_t ref = rxq->grant_ref[i];
314 
315 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
316 	rxq->grant_ref[i] = GRANT_REF_INVALID;
317 	return (ref);
318 }
319 
320 #define IPRINTK(fmt, args...) \
321     printf("[XEN] " fmt, ##args)
322 #ifdef INVARIANTS
323 #define WPRINTK(fmt, args...) \
324     printf("[XEN] " fmt, ##args)
325 #else
326 #define WPRINTK(fmt, args...)
327 #endif
328 #ifdef DEBUG
329 #define DPRINTK(fmt, args...) \
330     printf("[XEN] %s: " fmt, __func__, ##args)
331 #else
332 #define DPRINTK(fmt, args...)
333 #endif
334 
335 /**
336  * Read the 'mac' node at the given device's node in the store, and parse that
337  * as colon-separated octets, placing result the given mac array.  mac must be
338  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
339  * Return 0 on success, or errno on error.
340  */
341 static int
342 xen_net_read_mac(device_t dev, uint8_t mac[])
343 {
344 	int error, i;
345 	char *s, *e, *macstr;
346 	const char *path;
347 
348 	path = xenbus_get_node(dev);
349 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
350 	if (error == ENOENT) {
351 		/*
352 		 * Deal with missing mac XenStore nodes on devices with
353 		 * HVM emulation (the 'ioemu' configuration attribute)
354 		 * enabled.
355 		 *
356 		 * The HVM emulator may execute in a stub device model
357 		 * domain which lacks the permission, only given to Dom0,
358 		 * to update the guest's XenStore tree.  For this reason,
359 		 * the HVM emulator doesn't even attempt to write the
360 		 * front-side mac node, even when operating in Dom0.
361 		 * However, there should always be a mac listed in the
362 		 * backend tree.  Fallback to this version if our query
363 		 * of the front side XenStore location doesn't find
364 		 * anything.
365 		 */
366 		path = xenbus_get_otherend_path(dev);
367 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
368 	}
369 	if (error != 0) {
370 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
371 		return (error);
372 	}
373 
374 	s = macstr;
375 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
376 		mac[i] = strtoul(s, &e, 16);
377 		if (s == e || (e[0] != ':' && e[0] != 0)) {
378 			free(macstr, M_XENBUS);
379 			return (ENOENT);
380 		}
381 		s = &e[1];
382 	}
383 	free(macstr, M_XENBUS);
384 	return (0);
385 }
386 
387 /**
388  * Entry point to this code when a new device is created.  Allocate the basic
389  * structures and the ring buffers for communication with the backend, and
390  * inform the backend of the appropriate details for those.  Switch to
391  * Connected state.
392  */
393 static int
394 netfront_probe(device_t dev)
395 {
396 
397 	if (xen_hvm_domain() && xen_disable_pv_nics != 0)
398 		return (ENXIO);
399 
400 	if (!strcmp(xenbus_get_type(dev), "vif")) {
401 		device_set_desc(dev, "Virtual Network Interface");
402 		return (0);
403 	}
404 
405 	return (ENXIO);
406 }
407 
408 static int
409 netfront_attach(device_t dev)
410 {
411 	int err;
412 
413 	err = create_netdev(dev);
414 	if (err != 0) {
415 		xenbus_dev_fatal(dev, err, "creating netdev");
416 		return (err);
417 	}
418 
419 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
420 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
421 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
422 	    &xn_enable_lro, 0, "Large Receive Offload");
423 
424 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
425 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
426 	    OID_AUTO, "num_queues", CTLFLAG_RD,
427 	    &xn_num_queues, "Number of pairs of queues");
428 
429 	return (0);
430 }
431 
432 static int
433 netfront_suspend(device_t dev)
434 {
435 	struct netfront_info *np = device_get_softc(dev);
436 	u_int i;
437 
438 	for (i = 0; i < np->num_queues; i++) {
439 		XN_RX_LOCK(&np->rxq[i]);
440 		XN_TX_LOCK(&np->txq[i]);
441 	}
442 	netfront_carrier_off(np);
443 	for (i = 0; i < np->num_queues; i++) {
444 		XN_RX_UNLOCK(&np->rxq[i]);
445 		XN_TX_UNLOCK(&np->txq[i]);
446 	}
447 	return (0);
448 }
449 
450 /**
451  * We are reconnecting to the backend, due to a suspend/resume, or a backend
452  * driver restart.  We tear down our netif structure and recreate it, but
453  * leave the device-layer structures intact so that this is transparent to the
454  * rest of the kernel.
455  */
456 static int
457 netfront_resume(device_t dev)
458 {
459 	struct netfront_info *info = device_get_softc(dev);
460 
461 	info->xn_resume = true;
462 	netif_disconnect_backend(info);
463 	return (0);
464 }
465 
466 static int
467 write_queue_xenstore_keys(device_t dev,
468     struct netfront_rxq *rxq,
469     struct netfront_txq *txq,
470     struct xs_transaction *xst, bool hierarchy)
471 {
472 	int err;
473 	const char *message;
474 	const char *node = xenbus_get_node(dev);
475 	char *path;
476 	size_t path_size;
477 
478 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
479 	/* Split event channel support is not yet there. */
480 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
481 	    ("Split event channels are not supported"));
482 
483 	if (hierarchy) {
484 		path_size = strlen(node) + 10;
485 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
486 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
487 	} else {
488 		path_size = strlen(node) + 1;
489 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
490 		snprintf(path, path_size, "%s", node);
491 	}
492 
493 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
494 	if (err != 0) {
495 		message = "writing tx ring-ref";
496 		goto error;
497 	}
498 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
499 	if (err != 0) {
500 		message = "writing rx ring-ref";
501 		goto error;
502 	}
503 	err = xs_printf(*xst, path, "event-channel", "%u",
504 	    xen_intr_port(rxq->xen_intr_handle));
505 	if (err != 0) {
506 		message = "writing event-channel";
507 		goto error;
508 	}
509 
510 	free(path, M_DEVBUF);
511 
512 	return (0);
513 
514 error:
515 	free(path, M_DEVBUF);
516 	xenbus_dev_fatal(dev, err, "%s", message);
517 
518 	return (err);
519 }
520 
521 /* Common code used when first setting up, and when resuming. */
522 static int
523 talk_to_backend(device_t dev, struct netfront_info *info)
524 {
525 	const char *message;
526 	struct xs_transaction xst;
527 	const char *node = xenbus_get_node(dev);
528 	int err;
529 	unsigned long num_queues, max_queues = 0;
530 	unsigned int i;
531 
532 	err = xen_net_read_mac(dev, info->mac);
533 	if (err != 0) {
534 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
535 		goto out;
536 	}
537 
538 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
539 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
540 	if (err != 0)
541 		max_queues = 1;
542 	num_queues = xn_num_queues;
543 	if (num_queues > max_queues)
544 		num_queues = max_queues;
545 
546 	err = setup_device(dev, info, num_queues);
547 	if (err != 0)
548 		goto out;
549 
550  again:
551 	err = xs_transaction_start(&xst);
552 	if (err != 0) {
553 		xenbus_dev_fatal(dev, err, "starting transaction");
554 		goto free;
555 	}
556 
557 	if (info->num_queues == 1) {
558 		err = write_queue_xenstore_keys(dev, &info->rxq[0],
559 		    &info->txq[0], &xst, false);
560 		if (err != 0)
561 			goto abort_transaction_no_def_error;
562 	} else {
563 		err = xs_printf(xst, node, "multi-queue-num-queues",
564 		    "%u", info->num_queues);
565 		if (err != 0) {
566 			message = "writing multi-queue-num-queues";
567 			goto abort_transaction;
568 		}
569 
570 		for (i = 0; i < info->num_queues; i++) {
571 			err = write_queue_xenstore_keys(dev, &info->rxq[i],
572 			    &info->txq[i], &xst, true);
573 			if (err != 0)
574 				goto abort_transaction_no_def_error;
575 		}
576 	}
577 
578 	err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
579 	if (err != 0) {
580 		message = "writing request-rx-copy";
581 		goto abort_transaction;
582 	}
583 	err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
584 	if (err != 0) {
585 		message = "writing feature-rx-notify";
586 		goto abort_transaction;
587 	}
588 	err = xs_printf(xst, node, "feature-sg", "%d", 1);
589 	if (err != 0) {
590 		message = "writing feature-sg";
591 		goto abort_transaction;
592 	}
593 	err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
594 	if (err != 0) {
595 		message = "writing feature-gso-tcpv4";
596 		goto abort_transaction;
597 	}
598 
599 	err = xs_transaction_end(xst, 0);
600 	if (err != 0) {
601 		if (err == EAGAIN)
602 			goto again;
603 		xenbus_dev_fatal(dev, err, "completing transaction");
604 		goto free;
605 	}
606 
607 	return 0;
608 
609  abort_transaction:
610 	xenbus_dev_fatal(dev, err, "%s", message);
611  abort_transaction_no_def_error:
612 	xs_transaction_end(xst, 1);
613  free:
614 	netif_free(info);
615  out:
616 	return (err);
617 }
618 
619 static void
620 xn_rxq_intr(struct netfront_rxq *rxq)
621 {
622 
623 	XN_RX_LOCK(rxq);
624 	xn_rxeof(rxq);
625 	XN_RX_UNLOCK(rxq);
626 }
627 
628 static void
629 xn_txq_start(struct netfront_txq *txq)
630 {
631 	struct netfront_info *np = txq->info;
632 	struct ifnet *ifp = np->xn_ifp;
633 
634 	XN_TX_LOCK_ASSERT(txq);
635 	if (!drbr_empty(ifp, txq->br))
636 		xn_txq_mq_start_locked(txq, NULL);
637 }
638 
639 static void
640 xn_txq_intr(struct netfront_txq *txq)
641 {
642 
643 	XN_TX_LOCK(txq);
644 	if (RING_HAS_UNCONSUMED_RESPONSES(&txq->ring))
645 		xn_txeof(txq);
646 	xn_txq_start(txq);
647 	XN_TX_UNLOCK(txq);
648 }
649 
650 static void
651 xn_txq_tq_deferred(void *xtxq, int pending)
652 {
653 	struct netfront_txq *txq = xtxq;
654 
655 	XN_TX_LOCK(txq);
656 	xn_txq_start(txq);
657 	XN_TX_UNLOCK(txq);
658 }
659 
660 static void
661 disconnect_rxq(struct netfront_rxq *rxq)
662 {
663 
664 	xn_release_rx_bufs(rxq);
665 	gnttab_free_grant_references(rxq->gref_head);
666 	gnttab_end_foreign_access(rxq->ring_ref, NULL);
667 	/*
668 	 * No split event channel support at the moment, handle will
669 	 * be unbound in tx. So no need to call xen_intr_unbind here,
670 	 * but we do want to reset the handler to 0.
671 	 */
672 	rxq->xen_intr_handle = 0;
673 }
674 
675 static void
676 destroy_rxq(struct netfront_rxq *rxq)
677 {
678 
679 	callout_drain(&rxq->rx_refill);
680 	free(rxq->ring.sring, M_DEVBUF);
681 }
682 
683 static void
684 destroy_rxqs(struct netfront_info *np)
685 {
686 	int i;
687 
688 	for (i = 0; i < np->num_queues; i++)
689 		destroy_rxq(&np->rxq[i]);
690 
691 	free(np->rxq, M_DEVBUF);
692 	np->rxq = NULL;
693 }
694 
695 static int
696 setup_rxqs(device_t dev, struct netfront_info *info,
697 	   unsigned long num_queues)
698 {
699 	int q, i;
700 	int error;
701 	netif_rx_sring_t *rxs;
702 	struct netfront_rxq *rxq;
703 
704 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
705 	    M_DEVBUF, M_WAITOK|M_ZERO);
706 
707 	for (q = 0; q < num_queues; q++) {
708 		rxq = &info->rxq[q];
709 
710 		rxq->id = q;
711 		rxq->info = info;
712 		rxq->ring_ref = GRANT_REF_INVALID;
713 		rxq->ring.sring = NULL;
714 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
715 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
716 		    MTX_DEF);
717 
718 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
719 			rxq->mbufs[i] = NULL;
720 			rxq->grant_ref[i] = GRANT_REF_INVALID;
721 		}
722 
723 		/* Start resources allocation */
724 
725 		if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
726 		    &rxq->gref_head) != 0) {
727 			device_printf(dev, "allocating rx gref");
728 			error = ENOMEM;
729 			goto fail;
730 		}
731 
732 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
733 		    M_WAITOK|M_ZERO);
734 		SHARED_RING_INIT(rxs);
735 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
736 
737 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
738 		    &rxq->ring_ref);
739 		if (error != 0) {
740 			device_printf(dev, "granting rx ring page");
741 			goto fail_grant_ring;
742 		}
743 
744 		callout_init(&rxq->rx_refill, 1);
745 	}
746 
747 	return (0);
748 
749 fail_grant_ring:
750 	gnttab_free_grant_references(rxq->gref_head);
751 	free(rxq->ring.sring, M_DEVBUF);
752 fail:
753 	for (; q >= 0; q--) {
754 		disconnect_rxq(&info->rxq[q]);
755 		destroy_rxq(&info->rxq[q]);
756 	}
757 
758 	free(info->rxq, M_DEVBUF);
759 	return (error);
760 }
761 
762 static void
763 disconnect_txq(struct netfront_txq *txq)
764 {
765 
766 	xn_release_tx_bufs(txq);
767 	gnttab_free_grant_references(txq->gref_head);
768 	gnttab_end_foreign_access(txq->ring_ref, NULL);
769 	xen_intr_unbind(&txq->xen_intr_handle);
770 }
771 
772 static void
773 destroy_txq(struct netfront_txq *txq)
774 {
775 
776 	free(txq->ring.sring, M_DEVBUF);
777 	buf_ring_free(txq->br, M_DEVBUF);
778 	taskqueue_drain_all(txq->tq);
779 	taskqueue_free(txq->tq);
780 }
781 
782 static void
783 destroy_txqs(struct netfront_info *np)
784 {
785 	int i;
786 
787 	for (i = 0; i < np->num_queues; i++)
788 		destroy_txq(&np->txq[i]);
789 
790 	free(np->txq, M_DEVBUF);
791 	np->txq = NULL;
792 }
793 
794 static int
795 setup_txqs(device_t dev, struct netfront_info *info,
796 	   unsigned long num_queues)
797 {
798 	int q, i;
799 	int error;
800 	netif_tx_sring_t *txs;
801 	struct netfront_txq *txq;
802 
803 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
804 	    M_DEVBUF, M_WAITOK|M_ZERO);
805 
806 	for (q = 0; q < num_queues; q++) {
807 		txq = &info->txq[q];
808 
809 		txq->id = q;
810 		txq->info = info;
811 
812 		txq->ring_ref = GRANT_REF_INVALID;
813 		txq->ring.sring = NULL;
814 
815 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
816 
817 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
818 		    MTX_DEF);
819 
820 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
821 			txq->mbufs[i] = (void *) ((u_long) i+1);
822 			txq->grant_ref[i] = GRANT_REF_INVALID;
823 		}
824 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
825 
826 		/* Start resources allocation. */
827 
828 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
829 		    &txq->gref_head) != 0) {
830 			device_printf(dev, "failed to allocate tx grant refs\n");
831 			error = ENOMEM;
832 			goto fail;
833 		}
834 
835 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
836 		    M_WAITOK|M_ZERO);
837 		SHARED_RING_INIT(txs);
838 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
839 
840 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
841 		    &txq->ring_ref);
842 		if (error != 0) {
843 			device_printf(dev, "failed to grant tx ring\n");
844 			goto fail_grant_ring;
845 		}
846 
847 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
848 		    M_WAITOK, &txq->lock);
849 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
850 
851 		txq->tq = taskqueue_create(txq->name, M_WAITOK,
852 		    taskqueue_thread_enqueue, &txq->tq);
853 
854 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
855 		    "%s txq %d", device_get_nameunit(dev), txq->id);
856 		if (error != 0) {
857 			device_printf(dev, "failed to start tx taskq %d\n",
858 			    txq->id);
859 			goto fail_start_thread;
860 		}
861 
862 		error = xen_intr_alloc_and_bind_local_port(dev,
863 		    xenbus_get_otherend_id(dev), /* filter */ NULL, xn_intr,
864 		    &info->txq[q], INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
865 		    &txq->xen_intr_handle);
866 
867 		if (error != 0) {
868 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
869 			goto fail_bind_port;
870 		}
871 	}
872 
873 	return (0);
874 
875 fail_bind_port:
876 	taskqueue_drain_all(txq->tq);
877 fail_start_thread:
878 	buf_ring_free(txq->br, M_DEVBUF);
879 	taskqueue_free(txq->tq);
880 	gnttab_end_foreign_access(txq->ring_ref, NULL);
881 fail_grant_ring:
882 	gnttab_free_grant_references(txq->gref_head);
883 	free(txq->ring.sring, M_DEVBUF);
884 fail:
885 	for (; q >= 0; q--) {
886 		disconnect_txq(&info->txq[q]);
887 		destroy_txq(&info->txq[q]);
888 	}
889 
890 	free(info->txq, M_DEVBUF);
891 	return (error);
892 }
893 
894 static int
895 setup_device(device_t dev, struct netfront_info *info,
896     unsigned long num_queues)
897 {
898 	int error;
899 	int q;
900 
901 	if (info->txq)
902 		destroy_txqs(info);
903 
904 	if (info->rxq)
905 		destroy_rxqs(info);
906 
907 	info->num_queues = 0;
908 
909 	error = setup_rxqs(dev, info, num_queues);
910 	if (error != 0)
911 		goto out;
912 	error = setup_txqs(dev, info, num_queues);
913 	if (error != 0)
914 		goto out;
915 
916 	info->num_queues = num_queues;
917 
918 	/* No split event channel at the moment. */
919 	for (q = 0; q < num_queues; q++)
920 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
921 
922 	return (0);
923 
924 out:
925 	KASSERT(error != 0, ("Error path taken without providing an error code"));
926 	return (error);
927 }
928 
929 #ifdef INET
930 /**
931  * If this interface has an ipv4 address, send an arp for it. This
932  * helps to get the network going again after migrating hosts.
933  */
934 static void
935 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
936 {
937 	struct ifnet *ifp;
938 	struct ifaddr *ifa;
939 
940 	ifp = info->xn_ifp;
941 	TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
942 		if (ifa->ifa_addr->sa_family == AF_INET) {
943 			arp_ifinit(ifp, ifa);
944 		}
945 	}
946 }
947 #endif
948 
949 /**
950  * Callback received when the backend's state changes.
951  */
952 static void
953 netfront_backend_changed(device_t dev, XenbusState newstate)
954 {
955 	struct netfront_info *sc = device_get_softc(dev);
956 
957 	DPRINTK("newstate=%d\n", newstate);
958 
959 	switch (newstate) {
960 	case XenbusStateInitialising:
961 	case XenbusStateInitialised:
962 	case XenbusStateUnknown:
963 	case XenbusStateClosed:
964 	case XenbusStateReconfigured:
965 	case XenbusStateReconfiguring:
966 		break;
967 	case XenbusStateInitWait:
968 		if (xenbus_get_state(dev) != XenbusStateInitialising)
969 			break;
970 		if (xn_connect(sc) != 0)
971 			break;
972 		xenbus_set_state(dev, XenbusStateConnected);
973 		break;
974 	case XenbusStateClosing:
975 		xenbus_set_state(dev, XenbusStateClosed);
976 		break;
977 	case XenbusStateConnected:
978 #ifdef INET
979 		netfront_send_fake_arp(dev, sc);
980 #endif
981 		break;
982 	}
983 }
984 
985 /**
986  * \brief Verify that there is sufficient space in the Tx ring
987  *        buffer for a maximally sized request to be enqueued.
988  *
989  * A transmit request requires a transmit descriptor for each packet
990  * fragment, plus up to 2 entries for "options" (e.g. TSO).
991  */
992 static inline int
993 xn_tx_slot_available(struct netfront_txq *txq)
994 {
995 
996 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
997 }
998 
999 static void
1000 xn_release_tx_bufs(struct netfront_txq *txq)
1001 {
1002 	int i;
1003 
1004 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1005 		struct mbuf *m;
1006 
1007 		m = txq->mbufs[i];
1008 
1009 		/*
1010 		 * We assume that no kernel addresses are
1011 		 * less than NET_TX_RING_SIZE.  Any entry
1012 		 * in the table that is below this number
1013 		 * must be an index from free-list tracking.
1014 		 */
1015 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1016 			continue;
1017 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1018 		gnttab_release_grant_reference(&txq->gref_head,
1019 		    txq->grant_ref[i]);
1020 		txq->grant_ref[i] = GRANT_REF_INVALID;
1021 		add_id_to_freelist(txq->mbufs, i);
1022 		txq->mbufs_cnt--;
1023 		if (txq->mbufs_cnt < 0) {
1024 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1025 		}
1026 		m_free(m);
1027 	}
1028 }
1029 
1030 static struct mbuf *
1031 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1032 {
1033 	struct mbuf *m;
1034 
1035 	m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1036 	if (m == NULL)
1037 		return NULL;
1038 	m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1039 
1040 	return (m);
1041 }
1042 
1043 static void
1044 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1045 {
1046 	RING_IDX req_prod;
1047 	int notify;
1048 
1049 	XN_RX_LOCK_ASSERT(rxq);
1050 
1051 	if (__predict_false(rxq->info->carrier == 0))
1052 		return;
1053 
1054 	for (req_prod = rxq->ring.req_prod_pvt;
1055 	     req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1056 	     req_prod++) {
1057 		struct mbuf *m;
1058 		unsigned short id;
1059 		grant_ref_t ref;
1060 		struct netif_rx_request *req;
1061 		unsigned long pfn;
1062 
1063 		m = xn_alloc_one_rx_buffer(rxq);
1064 		if (m == NULL)
1065 			break;
1066 
1067 		id = xn_rxidx(req_prod);
1068 
1069 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1070 		rxq->mbufs[id] = m;
1071 
1072 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1073 		KASSERT(ref != GNTTAB_LIST_END,
1074 		    ("reserved grant references exhuasted"));
1075 		rxq->grant_ref[id] = ref;
1076 
1077 		pfn = atop(vtophys(mtod(m, vm_offset_t)));
1078 		req = RING_GET_REQUEST(&rxq->ring, req_prod);
1079 
1080 		gnttab_grant_foreign_access_ref(ref,
1081 		    xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1082 		req->id = id;
1083 		req->gref = ref;
1084 	}
1085 
1086 	rxq->ring.req_prod_pvt = req_prod;
1087 
1088 	/* Not enough requests? Try again later. */
1089 	if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1090 		callout_reset(&rxq->rx_refill, hz/10, xn_alloc_rx_buffers_callout,
1091 		    rxq);
1092 		return;
1093 	}
1094 
1095 	wmb();		/* barrier so backend seens requests */
1096 
1097 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1098 	if (notify)
1099 		xen_intr_signal(rxq->xen_intr_handle);
1100 }
1101 
1102 static void xn_alloc_rx_buffers_callout(void *arg)
1103 {
1104 	struct netfront_rxq *rxq;
1105 
1106 	rxq = (struct netfront_rxq *)arg;
1107 	XN_RX_LOCK(rxq);
1108 	xn_alloc_rx_buffers(rxq);
1109 	XN_RX_UNLOCK(rxq);
1110 }
1111 
1112 static void
1113 xn_release_rx_bufs(struct netfront_rxq *rxq)
1114 {
1115 	int i,  ref;
1116 	struct mbuf *m;
1117 
1118 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1119 		m = rxq->mbufs[i];
1120 
1121 		if (m == NULL)
1122 			continue;
1123 
1124 		ref = rxq->grant_ref[i];
1125 		if (ref == GRANT_REF_INVALID)
1126 			continue;
1127 
1128 		gnttab_end_foreign_access_ref(ref);
1129 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1130 		rxq->mbufs[i] = NULL;
1131 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1132 		m_freem(m);
1133 	}
1134 }
1135 
1136 static void
1137 xn_rxeof(struct netfront_rxq *rxq)
1138 {
1139 	struct ifnet *ifp;
1140 	struct netfront_info *np = rxq->info;
1141 #if (defined(INET) || defined(INET6))
1142 	struct lro_ctrl *lro = &rxq->lro;
1143 #endif
1144 	struct netfront_rx_info rinfo;
1145 	struct netif_rx_response *rx = &rinfo.rx;
1146 	struct netif_extra_info *extras = rinfo.extras;
1147 	RING_IDX i, rp;
1148 	struct mbuf *m;
1149 	struct mbufq mbufq_rxq, mbufq_errq;
1150 	int err, work_to_do;
1151 
1152 	do {
1153 		XN_RX_LOCK_ASSERT(rxq);
1154 		if (!netfront_carrier_ok(np))
1155 			return;
1156 
1157 		/* XXX: there should be some sane limit. */
1158 		mbufq_init(&mbufq_errq, INT_MAX);
1159 		mbufq_init(&mbufq_rxq, INT_MAX);
1160 
1161 		ifp = np->xn_ifp;
1162 
1163 		rp = rxq->ring.sring->rsp_prod;
1164 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1165 
1166 		i = rxq->ring.rsp_cons;
1167 		while ((i != rp)) {
1168 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1169 			memset(extras, 0, sizeof(rinfo.extras));
1170 
1171 			m = NULL;
1172 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1173 
1174 			if (__predict_false(err)) {
1175 				if (m)
1176 					(void )mbufq_enqueue(&mbufq_errq, m);
1177 				rxq->stats.rx_errors++;
1178 				continue;
1179 			}
1180 
1181 			m->m_pkthdr.rcvif = ifp;
1182 			if ( rx->flags & NETRXF_data_validated ) {
1183 				/* Tell the stack the checksums are okay */
1184 				/*
1185 				 * XXX this isn't necessarily the case - need to add
1186 				 * check
1187 				 */
1188 
1189 				m->m_pkthdr.csum_flags |=
1190 					(CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID
1191 					    | CSUM_PSEUDO_HDR);
1192 				m->m_pkthdr.csum_data = 0xffff;
1193 			}
1194 			if ((rx->flags & NETRXF_extra_info) != 0 &&
1195 			    (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1196 			    XEN_NETIF_EXTRA_TYPE_GSO)) {
1197 				m->m_pkthdr.tso_segsz =
1198 				extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1199 				m->m_pkthdr.csum_flags |= CSUM_TSO;
1200 			}
1201 
1202 			rxq->stats.rx_packets++;
1203 			rxq->stats.rx_bytes += m->m_pkthdr.len;
1204 
1205 			(void )mbufq_enqueue(&mbufq_rxq, m);
1206 			rxq->ring.rsp_cons = i;
1207 		}
1208 
1209 		mbufq_drain(&mbufq_errq);
1210 
1211 		/*
1212 		 * Process all the mbufs after the remapping is complete.
1213 		 * Break the mbuf chain first though.
1214 		 */
1215 		while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1216 			if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1217 
1218 			/* XXX: Do we really need to drop the rx lock? */
1219 			XN_RX_UNLOCK(rxq);
1220 #if (defined(INET) || defined(INET6))
1221 			/* Use LRO if possible */
1222 			if ((ifp->if_capenable & IFCAP_LRO) == 0 ||
1223 			    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1224 				/*
1225 				 * If LRO fails, pass up to the stack
1226 				 * directly.
1227 				 */
1228 				(*ifp->if_input)(ifp, m);
1229 			}
1230 #else
1231 			(*ifp->if_input)(ifp, m);
1232 #endif
1233 
1234 			XN_RX_LOCK(rxq);
1235 		}
1236 
1237 		rxq->ring.rsp_cons = i;
1238 
1239 #if (defined(INET) || defined(INET6))
1240 		/*
1241 		 * Flush any outstanding LRO work
1242 		 */
1243 		tcp_lro_flush_all(lro);
1244 #endif
1245 
1246 		xn_alloc_rx_buffers(rxq);
1247 
1248 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1249 	} while (work_to_do);
1250 }
1251 
1252 static void
1253 xn_txeof(struct netfront_txq *txq)
1254 {
1255 	RING_IDX i, prod;
1256 	unsigned short id;
1257 	struct ifnet *ifp;
1258 	netif_tx_response_t *txr;
1259 	struct mbuf *m;
1260 	struct netfront_info *np = txq->info;
1261 
1262 	XN_TX_LOCK_ASSERT(txq);
1263 
1264 	if (!netfront_carrier_ok(np))
1265 		return;
1266 
1267 	ifp = np->xn_ifp;
1268 
1269 	do {
1270 		prod = txq->ring.sring->rsp_prod;
1271 		rmb(); /* Ensure we see responses up to 'rp'. */
1272 
1273 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1274 			txr = RING_GET_RESPONSE(&txq->ring, i);
1275 			if (txr->status == NETIF_RSP_NULL)
1276 				continue;
1277 
1278 			if (txr->status != NETIF_RSP_OKAY) {
1279 				printf("%s: WARNING: response is %d!\n",
1280 				       __func__, txr->status);
1281 			}
1282 			id = txr->id;
1283 			m = txq->mbufs[id];
1284 			KASSERT(m != NULL, ("mbuf not found in chain"));
1285 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1286 				("mbuf already on the free list, but we're "
1287 				"trying to free it again!"));
1288 			M_ASSERTVALID(m);
1289 
1290 			/*
1291 			 * Increment packet count if this is the last
1292 			 * mbuf of the chain.
1293 			 */
1294 			if (!m->m_next)
1295 				if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1296 			if (__predict_false(gnttab_query_foreign_access(
1297 			    txq->grant_ref[id]) != 0)) {
1298 				panic("%s: grant id %u still in use by the "
1299 				    "backend", __func__, id);
1300 			}
1301 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1302 			gnttab_release_grant_reference(
1303 				&txq->gref_head, txq->grant_ref[id]);
1304 			txq->grant_ref[id] = GRANT_REF_INVALID;
1305 
1306 			txq->mbufs[id] = NULL;
1307 			add_id_to_freelist(txq->mbufs, id);
1308 			txq->mbufs_cnt--;
1309 			m_free(m);
1310 			/* Only mark the txq active if we've freed up at least one slot to try */
1311 			ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1312 		}
1313 		txq->ring.rsp_cons = prod;
1314 
1315 		/*
1316 		 * Set a new event, then check for race with update of
1317 		 * tx_cons. Note that it is essential to schedule a
1318 		 * callback, no matter how few buffers are pending. Even if
1319 		 * there is space in the transmit ring, higher layers may
1320 		 * be blocked because too much data is outstanding: in such
1321 		 * cases notification from Xen is likely to be the only kick
1322 		 * that we'll get.
1323 		 */
1324 		txq->ring.sring->rsp_event =
1325 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1326 
1327 		mb();
1328 	} while (prod != txq->ring.sring->rsp_prod);
1329 
1330 	if (txq->full &&
1331 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1332 		txq->full = false;
1333 		xn_txq_start(txq);
1334 	}
1335 }
1336 
1337 static void
1338 xn_intr(void *xsc)
1339 {
1340 	struct netfront_txq *txq = xsc;
1341 	struct netfront_info *np = txq->info;
1342 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1343 
1344 	/* kick both tx and rx */
1345 	xn_rxq_intr(rxq);
1346 	xn_txq_intr(txq);
1347 }
1348 
1349 static void
1350 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1351     grant_ref_t ref)
1352 {
1353 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1354 
1355 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1356 	rxq->mbufs[new] = m;
1357 	rxq->grant_ref[new] = ref;
1358 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1359 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1360 	rxq->ring.req_prod_pvt++;
1361 }
1362 
1363 static int
1364 xn_get_extras(struct netfront_rxq *rxq,
1365     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1366 {
1367 	struct netif_extra_info *extra;
1368 
1369 	int err = 0;
1370 
1371 	do {
1372 		struct mbuf *m;
1373 		grant_ref_t ref;
1374 
1375 		if (__predict_false(*cons + 1 == rp)) {
1376 			err = EINVAL;
1377 			break;
1378 		}
1379 
1380 		extra = (struct netif_extra_info *)
1381 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1382 
1383 		if (__predict_false(!extra->type ||
1384 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1385 			err = EINVAL;
1386 		} else {
1387 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1388 		}
1389 
1390 		m = xn_get_rx_mbuf(rxq, *cons);
1391 		ref = xn_get_rx_ref(rxq,  *cons);
1392 		xn_move_rx_slot(rxq, m, ref);
1393 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1394 
1395 	return err;
1396 }
1397 
1398 static int
1399 xn_get_responses(struct netfront_rxq *rxq,
1400     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1401     struct mbuf  **list)
1402 {
1403 	struct netif_rx_response *rx = &rinfo->rx;
1404 	struct netif_extra_info *extras = rinfo->extras;
1405 	struct mbuf *m, *m0, *m_prev;
1406 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1407 	RING_IDX ref_cons = *cons;
1408 	int frags = 1;
1409 	int err = 0;
1410 	u_long ret;
1411 
1412 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1413 
1414 	if (rx->flags & NETRXF_extra_info) {
1415 		err = xn_get_extras(rxq, extras, rp, cons);
1416 	}
1417 
1418 	if (m0 != NULL) {
1419 		m0->m_pkthdr.len = 0;
1420 		m0->m_next = NULL;
1421 	}
1422 
1423 	for (;;) {
1424 #if 0
1425 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1426 			rx->status, rx->offset, frags);
1427 #endif
1428 		if (__predict_false(rx->status < 0 ||
1429 			rx->offset + rx->status > PAGE_SIZE)) {
1430 
1431 			xn_move_rx_slot(rxq, m, ref);
1432 			if (m0 == m)
1433 				m0 = NULL;
1434 			m = NULL;
1435 			err = EINVAL;
1436 			goto next_skip_queue;
1437 		}
1438 
1439 		/*
1440 		 * This definitely indicates a bug, either in this driver or in
1441 		 * the backend driver. In future this should flag the bad
1442 		 * situation to the system controller to reboot the backed.
1443 		 */
1444 		if (ref == GRANT_REF_INVALID) {
1445 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1446 			err = EINVAL;
1447 			goto next;
1448 		}
1449 
1450 		ret = gnttab_end_foreign_access_ref(ref);
1451 		KASSERT(ret, ("Unable to end access to grant references"));
1452 
1453 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1454 
1455 next:
1456 		if (m == NULL)
1457 			break;
1458 
1459 		m->m_len = rx->status;
1460 		m->m_data += rx->offset;
1461 		m0->m_pkthdr.len += rx->status;
1462 
1463 next_skip_queue:
1464 		if (!(rx->flags & NETRXF_more_data))
1465 			break;
1466 
1467 		if (*cons + frags == rp) {
1468 			if (net_ratelimit())
1469 				WPRINTK("Need more frags\n");
1470 			err = ENOENT;
1471 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1472 			       __func__, *cons, frags, rp);
1473 			break;
1474 		}
1475 		/*
1476 		 * Note that m can be NULL, if rx->status < 0 or if
1477 		 * rx->offset + rx->status > PAGE_SIZE above.
1478 		 */
1479 		m_prev = m;
1480 
1481 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1482 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1483 
1484 		/*
1485 		 * m_prev == NULL can happen if rx->status < 0 or if
1486 		 * rx->offset + * rx->status > PAGE_SIZE above.
1487 		 */
1488 		if (m_prev != NULL)
1489 			m_prev->m_next = m;
1490 
1491 		/*
1492 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1493 		 * rx->status > PAGE_SIZE above.
1494 		 */
1495 		if (m0 == NULL)
1496 			m0 = m;
1497 		m->m_next = NULL;
1498 		ref = xn_get_rx_ref(rxq, *cons + frags);
1499 		ref_cons = *cons + frags;
1500 		frags++;
1501 	}
1502 	*list = m0;
1503 	*cons += frags;
1504 
1505 	return (err);
1506 }
1507 
1508 /**
1509  * \brief Count the number of fragments in an mbuf chain.
1510  *
1511  * Surprisingly, there isn't an M* macro for this.
1512  */
1513 static inline int
1514 xn_count_frags(struct mbuf *m)
1515 {
1516 	int nfrags;
1517 
1518 	for (nfrags = 0; m != NULL; m = m->m_next)
1519 		nfrags++;
1520 
1521 	return (nfrags);
1522 }
1523 
1524 /**
1525  * Given an mbuf chain, make sure we have enough room and then push
1526  * it onto the transmit ring.
1527  */
1528 static int
1529 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1530 {
1531 	struct mbuf *m;
1532 	struct netfront_info *np = txq->info;
1533 	struct ifnet *ifp = np->xn_ifp;
1534 	u_int nfrags;
1535 	int otherend_id;
1536 
1537 	/**
1538 	 * Defragment the mbuf if necessary.
1539 	 */
1540 	nfrags = xn_count_frags(m_head);
1541 
1542 	/*
1543 	 * Check to see whether this request is longer than netback
1544 	 * can handle, and try to defrag it.
1545 	 */
1546 	/**
1547 	 * It is a bit lame, but the netback driver in Linux can't
1548 	 * deal with nfrags > MAX_TX_REQ_FRAGS, which is a quirk of
1549 	 * the Linux network stack.
1550 	 */
1551 	if (nfrags > np->maxfrags) {
1552 		m = m_defrag(m_head, M_NOWAIT);
1553 		if (!m) {
1554 			/*
1555 			 * Defrag failed, so free the mbuf and
1556 			 * therefore drop the packet.
1557 			 */
1558 			m_freem(m_head);
1559 			return (EMSGSIZE);
1560 		}
1561 		m_head = m;
1562 	}
1563 
1564 	/* Determine how many fragments now exist */
1565 	nfrags = xn_count_frags(m_head);
1566 
1567 	/*
1568 	 * Check to see whether the defragmented packet has too many
1569 	 * segments for the Linux netback driver.
1570 	 */
1571 	/**
1572 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1573 	 * of mbufs longer than Linux can handle.  Make sure we don't
1574 	 * pass a too-long chain over to the other side by dropping the
1575 	 * packet.  It doesn't look like there is currently a way to
1576 	 * tell the TCP stack to generate a shorter chain of packets.
1577 	 */
1578 	if (nfrags > MAX_TX_REQ_FRAGS) {
1579 #ifdef DEBUG
1580 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1581 		       "won't be able to handle it, dropping\n",
1582 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1583 #endif
1584 		m_freem(m_head);
1585 		return (EMSGSIZE);
1586 	}
1587 
1588 	/*
1589 	 * This check should be redundant.  We've already verified that we
1590 	 * have enough slots in the ring to handle a packet of maximum
1591 	 * size, and that our packet is less than the maximum size.  Keep
1592 	 * it in here as an assert for now just to make certain that
1593 	 * chain_cnt is accurate.
1594 	 */
1595 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1596 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1597 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1598                     (int) nfrags, (int) NET_TX_RING_SIZE));
1599 
1600 	/*
1601 	 * Start packing the mbufs in this chain into
1602 	 * the fragment pointers. Stop when we run out
1603 	 * of fragments or hit the end of the mbuf chain.
1604 	 */
1605 	m = m_head;
1606 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1607 	for (m = m_head; m; m = m->m_next) {
1608 		netif_tx_request_t *tx;
1609 		uintptr_t id;
1610 		grant_ref_t ref;
1611 		u_long mfn; /* XXX Wrong type? */
1612 
1613 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1614 		id = get_id_from_freelist(txq->mbufs);
1615 		if (id == 0)
1616 			panic("%s: was allocated the freelist head!\n",
1617 			    __func__);
1618 		txq->mbufs_cnt++;
1619 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1620 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1621 			    __func__);
1622 		txq->mbufs[id] = m;
1623 		tx->id = id;
1624 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1625 		KASSERT((short)ref >= 0, ("Negative ref"));
1626 		mfn = virt_to_mfn(mtod(m, vm_offset_t));
1627 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1628 		    mfn, GNTMAP_readonly);
1629 		tx->gref = txq->grant_ref[id] = ref;
1630 		tx->offset = mtod(m, vm_offset_t) & (PAGE_SIZE - 1);
1631 		tx->flags = 0;
1632 		if (m == m_head) {
1633 			/*
1634 			 * The first fragment has the entire packet
1635 			 * size, subsequent fragments have just the
1636 			 * fragment size. The backend works out the
1637 			 * true size of the first fragment by
1638 			 * subtracting the sizes of the other
1639 			 * fragments.
1640 			 */
1641 			tx->size = m->m_pkthdr.len;
1642 
1643 			/*
1644 			 * The first fragment contains the checksum flags
1645 			 * and is optionally followed by extra data for
1646 			 * TSO etc.
1647 			 */
1648 			/**
1649 			 * CSUM_TSO requires checksum offloading.
1650 			 * Some versions of FreeBSD fail to
1651 			 * set CSUM_TCP in the CSUM_TSO case,
1652 			 * so we have to test for CSUM_TSO
1653 			 * explicitly.
1654 			 */
1655 			if (m->m_pkthdr.csum_flags
1656 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1657 				tx->flags |= (NETTXF_csum_blank
1658 				    | NETTXF_data_validated);
1659 			}
1660 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1661 				struct netif_extra_info *gso =
1662 					(struct netif_extra_info *)
1663 					RING_GET_REQUEST(&txq->ring,
1664 							 ++txq->ring.req_prod_pvt);
1665 
1666 				tx->flags |= NETTXF_extra_info;
1667 
1668 				gso->u.gso.size = m->m_pkthdr.tso_segsz;
1669 				gso->u.gso.type =
1670 					XEN_NETIF_GSO_TYPE_TCPV4;
1671 				gso->u.gso.pad = 0;
1672 				gso->u.gso.features = 0;
1673 
1674 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1675 				gso->flags = 0;
1676 			}
1677 		} else {
1678 			tx->size = m->m_len;
1679 		}
1680 		if (m->m_next)
1681 			tx->flags |= NETTXF_more_data;
1682 
1683 		txq->ring.req_prod_pvt++;
1684 	}
1685 	BPF_MTAP(ifp, m_head);
1686 
1687 	xn_txeof(txq);
1688 
1689 	txq->stats.tx_bytes += m_head->m_pkthdr.len;
1690 	txq->stats.tx_packets++;
1691 
1692 	return (0);
1693 }
1694 
1695 /* equivalent of network_open() in Linux */
1696 static void
1697 xn_ifinit_locked(struct netfront_info *np)
1698 {
1699 	struct ifnet *ifp;
1700 	int i;
1701 	struct netfront_rxq *rxq;
1702 
1703 	XN_LOCK_ASSERT(np);
1704 
1705 	ifp = np->xn_ifp;
1706 
1707 	if (ifp->if_drv_flags & IFF_DRV_RUNNING || !netfront_carrier_ok(np))
1708 		return;
1709 
1710 	xn_stop(np);
1711 
1712 	for (i = 0; i < np->num_queues; i++) {
1713 		rxq = &np->rxq[i];
1714 		XN_RX_LOCK(rxq);
1715 		xn_alloc_rx_buffers(rxq);
1716 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1717 		if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1718 			xn_rxeof(rxq);
1719 		XN_RX_UNLOCK(rxq);
1720 	}
1721 
1722 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
1723 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1724 	if_link_state_change(ifp, LINK_STATE_UP);
1725 }
1726 
1727 static void
1728 xn_ifinit(void *xsc)
1729 {
1730 	struct netfront_info *sc = xsc;
1731 
1732 	XN_LOCK(sc);
1733 	xn_ifinit_locked(sc);
1734 	XN_UNLOCK(sc);
1735 }
1736 
1737 static int
1738 xn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
1739 {
1740 	struct netfront_info *sc = ifp->if_softc;
1741 	struct ifreq *ifr = (struct ifreq *) data;
1742 #ifdef INET
1743 	struct ifaddr *ifa = (struct ifaddr *)data;
1744 #endif
1745 
1746 	int mask, error = 0;
1747 	switch(cmd) {
1748 	case SIOCSIFADDR:
1749 #ifdef INET
1750 		XN_LOCK(sc);
1751 		if (ifa->ifa_addr->sa_family == AF_INET) {
1752 			ifp->if_flags |= IFF_UP;
1753 			if (!(ifp->if_drv_flags & IFF_DRV_RUNNING))
1754 				xn_ifinit_locked(sc);
1755 			arp_ifinit(ifp, ifa);
1756 			XN_UNLOCK(sc);
1757 		} else {
1758 			XN_UNLOCK(sc);
1759 #endif
1760 			error = ether_ioctl(ifp, cmd, data);
1761 #ifdef INET
1762 		}
1763 #endif
1764 		break;
1765 	case SIOCSIFMTU:
1766 		ifp->if_mtu = ifr->ifr_mtu;
1767 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1768 		xn_ifinit(sc);
1769 		break;
1770 	case SIOCSIFFLAGS:
1771 		XN_LOCK(sc);
1772 		if (ifp->if_flags & IFF_UP) {
1773 			/*
1774 			 * If only the state of the PROMISC flag changed,
1775 			 * then just use the 'set promisc mode' command
1776 			 * instead of reinitializing the entire NIC. Doing
1777 			 * a full re-init means reloading the firmware and
1778 			 * waiting for it to start up, which may take a
1779 			 * second or two.
1780 			 */
1781 			xn_ifinit_locked(sc);
1782 		} else {
1783 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1784 				xn_stop(sc);
1785 			}
1786 		}
1787 		sc->xn_if_flags = ifp->if_flags;
1788 		XN_UNLOCK(sc);
1789 		break;
1790 	case SIOCSIFCAP:
1791 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
1792 		if (mask & IFCAP_TXCSUM) {
1793 			if (IFCAP_TXCSUM & ifp->if_capenable) {
1794 				ifp->if_capenable &= ~(IFCAP_TXCSUM|IFCAP_TSO4);
1795 				ifp->if_hwassist &= ~(CSUM_TCP | CSUM_UDP
1796 				    | CSUM_IP | CSUM_TSO);
1797 			} else {
1798 				ifp->if_capenable |= IFCAP_TXCSUM;
1799 				ifp->if_hwassist |= (CSUM_TCP | CSUM_UDP
1800 				    | CSUM_IP);
1801 			}
1802 		}
1803 		if (mask & IFCAP_RXCSUM) {
1804 			ifp->if_capenable ^= IFCAP_RXCSUM;
1805 		}
1806 		if (mask & IFCAP_TSO4) {
1807 			if (IFCAP_TSO4 & ifp->if_capenable) {
1808 				ifp->if_capenable &= ~IFCAP_TSO4;
1809 				ifp->if_hwassist &= ~CSUM_TSO;
1810 			} else if (IFCAP_TXCSUM & ifp->if_capenable) {
1811 				ifp->if_capenable |= IFCAP_TSO4;
1812 				ifp->if_hwassist |= CSUM_TSO;
1813 			} else {
1814 				IPRINTK("Xen requires tx checksum offload"
1815 				    " be enabled to use TSO\n");
1816 				error = EINVAL;
1817 			}
1818 		}
1819 		if (mask & IFCAP_LRO) {
1820 			ifp->if_capenable ^= IFCAP_LRO;
1821 
1822 		}
1823 		break;
1824 	case SIOCADDMULTI:
1825 	case SIOCDELMULTI:
1826 		break;
1827 	case SIOCSIFMEDIA:
1828 	case SIOCGIFMEDIA:
1829 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1830 		break;
1831 	default:
1832 		error = ether_ioctl(ifp, cmd, data);
1833 	}
1834 
1835 	return (error);
1836 }
1837 
1838 static void
1839 xn_stop(struct netfront_info *sc)
1840 {
1841 	struct ifnet *ifp;
1842 
1843 	XN_LOCK_ASSERT(sc);
1844 
1845 	ifp = sc->xn_ifp;
1846 
1847 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1848 	if_link_state_change(ifp, LINK_STATE_DOWN);
1849 }
1850 
1851 static void
1852 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1853 {
1854 	int requeue_idx, i;
1855 	grant_ref_t ref;
1856 	netif_rx_request_t *req;
1857 
1858 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1859 		struct mbuf *m;
1860 		u_long pfn;
1861 
1862 		if (rxq->mbufs[i] == NULL)
1863 			continue;
1864 
1865 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1866 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1867 
1868 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1869 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1870 
1871 		gnttab_grant_foreign_access_ref(ref,
1872 		    xenbus_get_otherend_id(rxq->info->xbdev),
1873 		    pfn, 0);
1874 
1875 		req->gref = ref;
1876 		req->id   = requeue_idx;
1877 
1878 		requeue_idx++;
1879 	}
1880 
1881 	rxq->ring.req_prod_pvt = requeue_idx;
1882 }
1883 
1884 /* START of Xenolinux helper functions adapted to FreeBSD */
1885 int
1886 xn_connect(struct netfront_info *np)
1887 {
1888 	int i, error;
1889 	u_int feature_rx_copy;
1890 	struct netfront_rxq *rxq;
1891 	struct netfront_txq *txq;
1892 
1893 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1894 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
1895 	if (error != 0)
1896 		feature_rx_copy = 0;
1897 
1898 	/* We only support rx copy. */
1899 	if (!feature_rx_copy)
1900 		return (EPROTONOSUPPORT);
1901 
1902 	/* Recovery procedure: */
1903 	error = talk_to_backend(np->xbdev, np);
1904 	if (error != 0)
1905 		return (error);
1906 
1907 	/* Step 1: Reinitialise variables. */
1908 	xn_query_features(np);
1909 	xn_configure_features(np);
1910 
1911 	/* Step 2: Release TX buffer */
1912 	for (i = 0; i < np->num_queues; i++) {
1913 		txq = &np->txq[i];
1914 		xn_release_tx_bufs(txq);
1915 	}
1916 
1917 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
1918 	for (i = 0; i < np->num_queues; i++) {
1919 		rxq = &np->rxq[i];
1920 		xn_rebuild_rx_bufs(rxq);
1921 	}
1922 
1923 	/* Step 4: All public and private state should now be sane.  Get
1924 	 * ready to start sending and receiving packets and give the driver
1925 	 * domain a kick because we've probably just requeued some
1926 	 * packets.
1927 	 */
1928 	netfront_carrier_on(np);
1929 	for (i = 0; i < np->num_queues; i++) {
1930 		txq = &np->txq[i];
1931 		xen_intr_signal(txq->xen_intr_handle);
1932 		XN_TX_LOCK(txq);
1933 		xn_txeof(txq);
1934 		XN_TX_UNLOCK(txq);
1935 		XN_RX_LOCK(rxq);
1936 		xn_alloc_rx_buffers(rxq);
1937 		XN_RX_UNLOCK(rxq);
1938 	}
1939 
1940 	return (0);
1941 }
1942 
1943 static void
1944 xn_query_features(struct netfront_info *np)
1945 {
1946 	int val;
1947 
1948 	device_printf(np->xbdev, "backend features:");
1949 
1950 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1951 		"feature-sg", NULL, "%d", &val) != 0)
1952 		val = 0;
1953 
1954 	np->maxfrags = 1;
1955 	if (val) {
1956 		np->maxfrags = MAX_TX_REQ_FRAGS;
1957 		printf(" feature-sg");
1958 	}
1959 
1960 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1961 		"feature-gso-tcpv4", NULL, "%d", &val) != 0)
1962 		val = 0;
1963 
1964 	np->xn_ifp->if_capabilities &= ~(IFCAP_TSO4|IFCAP_LRO);
1965 	if (val) {
1966 		np->xn_ifp->if_capabilities |= IFCAP_TSO4|IFCAP_LRO;
1967 		printf(" feature-gso-tcp4");
1968 	}
1969 
1970 	printf("\n");
1971 }
1972 
1973 static int
1974 xn_configure_features(struct netfront_info *np)
1975 {
1976 	int err, cap_enabled;
1977 #if (defined(INET) || defined(INET6))
1978 	int i;
1979 #endif
1980 
1981 	err = 0;
1982 
1983 	if (np->xn_resume &&
1984 	    ((np->xn_ifp->if_capenable & np->xn_ifp->if_capabilities)
1985 	    == np->xn_ifp->if_capenable)) {
1986 		/* Current options are available, no need to do anything. */
1987 		return (0);
1988 	}
1989 
1990 	/* Try to preserve as many options as possible. */
1991 	if (np->xn_resume)
1992 		cap_enabled = np->xn_ifp->if_capenable;
1993 	else
1994 		cap_enabled = UINT_MAX;
1995 
1996 #if (defined(INET) || defined(INET6))
1997 	for (i = 0; i < np->num_queues; i++)
1998 		if ((np->xn_ifp->if_capenable & IFCAP_LRO) ==
1999 		    (cap_enabled & IFCAP_LRO))
2000 			tcp_lro_free(&np->rxq[i].lro);
2001 #endif
2002     	np->xn_ifp->if_capenable =
2003 	    np->xn_ifp->if_capabilities & ~(IFCAP_LRO|IFCAP_TSO4) & cap_enabled;
2004 	np->xn_ifp->if_hwassist &= ~CSUM_TSO;
2005 #if (defined(INET) || defined(INET6))
2006 	for (i = 0; i < np->num_queues; i++) {
2007 		if (xn_enable_lro && (np->xn_ifp->if_capabilities & IFCAP_LRO) ==
2008 		    (cap_enabled & IFCAP_LRO)) {
2009 			err = tcp_lro_init(&np->rxq[i].lro);
2010 			if (err != 0) {
2011 				device_printf(np->xbdev, "LRO initialization failed\n");
2012 			} else {
2013 				np->rxq[i].lro.ifp = np->xn_ifp;
2014 				np->xn_ifp->if_capenable |= IFCAP_LRO;
2015 			}
2016 		}
2017 	}
2018 	if ((np->xn_ifp->if_capabilities & IFCAP_TSO4) ==
2019 	    (cap_enabled & IFCAP_TSO4)) {
2020 		np->xn_ifp->if_capenable |= IFCAP_TSO4;
2021 		np->xn_ifp->if_hwassist |= CSUM_TSO;
2022 	}
2023 #endif
2024 	return (err);
2025 }
2026 
2027 static int
2028 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2029 {
2030 	struct netfront_info *np;
2031 	struct ifnet *ifp;
2032 	struct buf_ring *br;
2033 	int error, notify;
2034 
2035 	np = txq->info;
2036 	br = txq->br;
2037 	ifp = np->xn_ifp;
2038 	error = 0;
2039 
2040 	XN_TX_LOCK_ASSERT(txq);
2041 
2042 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
2043 	    !netfront_carrier_ok(np)) {
2044 		if (m != NULL)
2045 			error = drbr_enqueue(ifp, br, m);
2046 		return (error);
2047 	}
2048 
2049 	if (m != NULL) {
2050 		error = drbr_enqueue(ifp, br, m);
2051 		if (error != 0)
2052 			return (error);
2053 	}
2054 
2055 	while ((m = drbr_peek(ifp, br)) != NULL) {
2056 		if (!xn_tx_slot_available(txq)) {
2057 			drbr_putback(ifp, br, m);
2058 			break;
2059 		}
2060 
2061 		error = xn_assemble_tx_request(txq, m);
2062 		/* xn_assemble_tx_request always consumes the mbuf*/
2063 		if (error != 0) {
2064 			drbr_advance(ifp, br);
2065 			break;
2066 		}
2067 
2068 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2069 		if (notify)
2070 			xen_intr_signal(txq->xen_intr_handle);
2071 
2072 		drbr_advance(ifp, br);
2073 	}
2074 
2075 	if (RING_FULL(&txq->ring))
2076 		txq->full = true;
2077 
2078 	return (0);
2079 }
2080 
2081 static int
2082 xn_txq_mq_start(struct ifnet *ifp, struct mbuf *m)
2083 {
2084 	struct netfront_info *np;
2085 	struct netfront_txq *txq;
2086 	int i, npairs, error;
2087 
2088 	np = ifp->if_softc;
2089 	npairs = np->num_queues;
2090 
2091 	KASSERT(npairs != 0, ("called with 0 available queues"));
2092 
2093 	/* check if flowid is set */
2094 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2095 		i = m->m_pkthdr.flowid % npairs;
2096 	else
2097 		i = curcpu % npairs;
2098 
2099 	txq = &np->txq[i];
2100 
2101 	if (XN_TX_TRYLOCK(txq) != 0) {
2102 		error = xn_txq_mq_start_locked(txq, m);
2103 		XN_TX_UNLOCK(txq);
2104 	} else {
2105 		error = drbr_enqueue(ifp, txq->br, m);
2106 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2107 	}
2108 
2109 	return (error);
2110 }
2111 
2112 static void
2113 xn_qflush(struct ifnet *ifp)
2114 {
2115 	struct netfront_info *np;
2116 	struct netfront_txq *txq;
2117 	struct mbuf *m;
2118 	int i;
2119 
2120 	np = ifp->if_softc;
2121 
2122 	for (i = 0; i < np->num_queues; i++) {
2123 		txq = &np->txq[i];
2124 
2125 		XN_TX_LOCK(txq);
2126 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2127 			m_freem(m);
2128 		XN_TX_UNLOCK(txq);
2129 	}
2130 
2131 	if_qflush(ifp);
2132 }
2133 
2134 /**
2135  * Create a network device.
2136  * @param dev  Newbus device representing this virtual NIC.
2137  */
2138 int
2139 create_netdev(device_t dev)
2140 {
2141 	struct netfront_info *np;
2142 	int err;
2143 	struct ifnet *ifp;
2144 
2145 	np = device_get_softc(dev);
2146 
2147 	np->xbdev         = dev;
2148 
2149 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2150 
2151 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2152 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2153 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2154 
2155 	err = xen_net_read_mac(dev, np->mac);
2156 	if (err != 0)
2157 		goto error;
2158 
2159 	/* Set up ifnet structure */
2160 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2161     	ifp->if_softc = np;
2162     	if_initname(ifp, "xn",  device_get_unit(dev));
2163     	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2164     	ifp->if_ioctl = xn_ioctl;
2165 
2166 	ifp->if_transmit = xn_txq_mq_start;
2167 	ifp->if_qflush = xn_qflush;
2168 
2169     	ifp->if_init = xn_ifinit;
2170 
2171     	ifp->if_hwassist = XN_CSUM_FEATURES;
2172     	ifp->if_capabilities = IFCAP_HWCSUM;
2173 	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2174 	ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
2175 	ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
2176 
2177     	ether_ifattach(ifp, np->mac);
2178 	netfront_carrier_off(np);
2179 
2180 	return (0);
2181 
2182 error:
2183 	KASSERT(err != 0, ("Error path with no error code specified"));
2184 	return (err);
2185 }
2186 
2187 static int
2188 netfront_detach(device_t dev)
2189 {
2190 	struct netfront_info *info = device_get_softc(dev);
2191 
2192 	DPRINTK("%s\n", xenbus_get_node(dev));
2193 
2194 	netif_free(info);
2195 
2196 	return 0;
2197 }
2198 
2199 static void
2200 netif_free(struct netfront_info *np)
2201 {
2202 
2203 	XN_LOCK(np);
2204 	xn_stop(np);
2205 	XN_UNLOCK(np);
2206 	netif_disconnect_backend(np);
2207 	ether_ifdetach(np->xn_ifp);
2208 	free(np->rxq, M_DEVBUF);
2209 	free(np->txq, M_DEVBUF);
2210 	if_free(np->xn_ifp);
2211 	np->xn_ifp = NULL;
2212 	ifmedia_removeall(&np->sc_media);
2213 }
2214 
2215 static void
2216 netif_disconnect_backend(struct netfront_info *np)
2217 {
2218 	u_int i;
2219 
2220 	for (i = 0; i < np->num_queues; i++) {
2221 		XN_RX_LOCK(&np->rxq[i]);
2222 		XN_TX_LOCK(&np->txq[i]);
2223 	}
2224 	netfront_carrier_off(np);
2225 	for (i = 0; i < np->num_queues; i++) {
2226 		XN_RX_UNLOCK(&np->rxq[i]);
2227 		XN_TX_UNLOCK(&np->txq[i]);
2228 	}
2229 
2230 	for (i = 0; i < np->num_queues; i++) {
2231 		disconnect_rxq(&np->rxq[i]);
2232 		disconnect_txq(&np->txq[i]);
2233 	}
2234 }
2235 
2236 static int
2237 xn_ifmedia_upd(struct ifnet *ifp)
2238 {
2239 
2240 	return (0);
2241 }
2242 
2243 static void
2244 xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
2245 {
2246 
2247 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2248 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2249 }
2250 
2251 /* ** Driver registration ** */
2252 static device_method_t netfront_methods[] = {
2253 	/* Device interface */
2254 	DEVMETHOD(device_probe,         netfront_probe),
2255 	DEVMETHOD(device_attach,        netfront_attach),
2256 	DEVMETHOD(device_detach,        netfront_detach),
2257 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2258 	DEVMETHOD(device_suspend,       netfront_suspend),
2259 	DEVMETHOD(device_resume,        netfront_resume),
2260 
2261 	/* Xenbus interface */
2262 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2263 
2264 	DEVMETHOD_END
2265 };
2266 
2267 static driver_t netfront_driver = {
2268 	"xn",
2269 	netfront_methods,
2270 	sizeof(struct netfront_info),
2271 };
2272 devclass_t netfront_devclass;
2273 
2274 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, netfront_devclass, NULL,
2275     NULL);
2276