xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision 2568ee67473c1dc634f5d6bcf9976987e052932c)
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(void *);
125 static void xn_txq_intr(void *);
126 static int  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 taskqueue 	*tq;
192 	struct task		intrtask;
193 
194 	struct callout		rx_refill;
195 
196 	struct xn_rx_stats	stats;
197 };
198 
199 struct netfront_txq {
200 	struct netfront_info 	*info;
201 	u_int 			id;
202 	char			name[XN_QUEUE_NAME_LEN];
203 	struct mtx		lock;
204 
205 	int			ring_ref;
206 	netif_tx_front_ring_t	ring;
207 	xen_intr_handle_t 	xen_intr_handle;
208 
209 	grant_ref_t		gref_head;
210 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
211 
212 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
213 	int			mbufs_cnt;
214 	struct buf_ring		*br;
215 
216 	struct taskqueue 	*tq;
217 	struct task       	intrtask;
218 	struct task       	defrtask;
219 
220 	bool			full;
221 
222 	struct xn_tx_stats	stats;
223 };
224 
225 struct netfront_info {
226 	struct ifnet 		*xn_ifp;
227 
228 	struct mtx   		sc_lock;
229 
230 	u_int  num_queues;
231 	struct netfront_rxq 	*rxq;
232 	struct netfront_txq 	*txq;
233 
234 	u_int			carrier;
235 	u_int			maxfrags;
236 
237 	device_t		xbdev;
238 	uint8_t			mac[ETHER_ADDR_LEN];
239 
240 	int			xn_if_flags;
241 
242 	struct ifmedia		sc_media;
243 
244 	bool			xn_resume;
245 };
246 
247 struct netfront_rx_info {
248 	struct netif_rx_response rx;
249 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
250 };
251 
252 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
253 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
254 
255 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
256 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
257 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
258 
259 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
260 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
261 
262 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
263 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
264 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
265 
266 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
267 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
268 #define netfront_carrier_ok(netif)	((netif)->carrier)
269 
270 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
271 
272 static inline void
273 add_id_to_freelist(struct mbuf **list, uintptr_t id)
274 {
275 
276 	KASSERT(id != 0,
277 		("%s: the head item (0) must always be free.", __func__));
278 	list[id] = list[0];
279 	list[0]  = (struct mbuf *)id;
280 }
281 
282 static inline unsigned short
283 get_id_from_freelist(struct mbuf **list)
284 {
285 	uintptr_t id;
286 
287 	id = (uintptr_t)list[0];
288 	KASSERT(id != 0,
289 		("%s: the head item (0) must always remain free.", __func__));
290 	list[0] = list[id];
291 	return (id);
292 }
293 
294 static inline int
295 xn_rxidx(RING_IDX idx)
296 {
297 
298 	return idx & (NET_RX_RING_SIZE - 1);
299 }
300 
301 static inline struct mbuf *
302 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
303 {
304 	int i;
305 	struct mbuf *m;
306 
307 	i = xn_rxidx(ri);
308 	m = rxq->mbufs[i];
309 	rxq->mbufs[i] = NULL;
310 	return (m);
311 }
312 
313 static inline grant_ref_t
314 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
315 {
316 	int i = xn_rxidx(ri);
317 	grant_ref_t ref = rxq->grant_ref[i];
318 
319 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
320 	rxq->grant_ref[i] = GRANT_REF_INVALID;
321 	return (ref);
322 }
323 
324 #define IPRINTK(fmt, args...) \
325     printf("[XEN] " fmt, ##args)
326 #ifdef INVARIANTS
327 #define WPRINTK(fmt, args...) \
328     printf("[XEN] " fmt, ##args)
329 #else
330 #define WPRINTK(fmt, args...)
331 #endif
332 #ifdef DEBUG
333 #define DPRINTK(fmt, args...) \
334     printf("[XEN] %s: " fmt, __func__, ##args)
335 #else
336 #define DPRINTK(fmt, args...)
337 #endif
338 
339 /**
340  * Read the 'mac' node at the given device's node in the store, and parse that
341  * as colon-separated octets, placing result the given mac array.  mac must be
342  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
343  * Return 0 on success, or errno on error.
344  */
345 static int
346 xen_net_read_mac(device_t dev, uint8_t mac[])
347 {
348 	int error, i;
349 	char *s, *e, *macstr;
350 	const char *path;
351 
352 	path = xenbus_get_node(dev);
353 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
354 	if (error == ENOENT) {
355 		/*
356 		 * Deal with missing mac XenStore nodes on devices with
357 		 * HVM emulation (the 'ioemu' configuration attribute)
358 		 * enabled.
359 		 *
360 		 * The HVM emulator may execute in a stub device model
361 		 * domain which lacks the permission, only given to Dom0,
362 		 * to update the guest's XenStore tree.  For this reason,
363 		 * the HVM emulator doesn't even attempt to write the
364 		 * front-side mac node, even when operating in Dom0.
365 		 * However, there should always be a mac listed in the
366 		 * backend tree.  Fallback to this version if our query
367 		 * of the front side XenStore location doesn't find
368 		 * anything.
369 		 */
370 		path = xenbus_get_otherend_path(dev);
371 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
372 	}
373 	if (error != 0) {
374 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
375 		return (error);
376 	}
377 
378 	s = macstr;
379 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
380 		mac[i] = strtoul(s, &e, 16);
381 		if (s == e || (e[0] != ':' && e[0] != 0)) {
382 			free(macstr, M_XENBUS);
383 			return (ENOENT);
384 		}
385 		s = &e[1];
386 	}
387 	free(macstr, M_XENBUS);
388 	return (0);
389 }
390 
391 /**
392  * Entry point to this code when a new device is created.  Allocate the basic
393  * structures and the ring buffers for communication with the backend, and
394  * inform the backend of the appropriate details for those.  Switch to
395  * Connected state.
396  */
397 static int
398 netfront_probe(device_t dev)
399 {
400 
401 	if (xen_hvm_domain() && xen_disable_pv_nics != 0)
402 		return (ENXIO);
403 
404 	if (!strcmp(xenbus_get_type(dev), "vif")) {
405 		device_set_desc(dev, "Virtual Network Interface");
406 		return (0);
407 	}
408 
409 	return (ENXIO);
410 }
411 
412 static int
413 netfront_attach(device_t dev)
414 {
415 	int err;
416 
417 	err = create_netdev(dev);
418 	if (err != 0) {
419 		xenbus_dev_fatal(dev, err, "creating netdev");
420 		return (err);
421 	}
422 
423 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
424 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
425 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
426 	    &xn_enable_lro, 0, "Large Receive Offload");
427 
428 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
429 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
430 	    OID_AUTO, "num_queues", CTLFLAG_RD,
431 	    &xn_num_queues, "Number of pairs of queues");
432 
433 	return (0);
434 }
435 
436 static int
437 netfront_suspend(device_t dev)
438 {
439 	struct netfront_info *np = device_get_softc(dev);
440 	u_int i;
441 
442 	for (i = 0; i < np->num_queues; i++) {
443 		XN_RX_LOCK(&np->rxq[i]);
444 		XN_TX_LOCK(&np->txq[i]);
445 	}
446 	netfront_carrier_off(np);
447 	for (i = 0; i < np->num_queues; i++) {
448 		XN_RX_UNLOCK(&np->rxq[i]);
449 		XN_TX_UNLOCK(&np->txq[i]);
450 	}
451 	return (0);
452 }
453 
454 /**
455  * We are reconnecting to the backend, due to a suspend/resume, or a backend
456  * driver restart.  We tear down our netif structure and recreate it, but
457  * leave the device-layer structures intact so that this is transparent to the
458  * rest of the kernel.
459  */
460 static int
461 netfront_resume(device_t dev)
462 {
463 	struct netfront_info *info = device_get_softc(dev);
464 
465 	info->xn_resume = true;
466 	netif_disconnect_backend(info);
467 	return (0);
468 }
469 
470 static int
471 write_queue_xenstore_keys(device_t dev,
472     struct netfront_rxq *rxq,
473     struct netfront_txq *txq,
474     struct xs_transaction *xst, bool hierarchy)
475 {
476 	int err;
477 	const char *message;
478 	const char *node = xenbus_get_node(dev);
479 	char *path;
480 	size_t path_size;
481 
482 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
483 	/* Split event channel support is not yet there. */
484 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
485 	    ("Split event channels are not supported"));
486 
487 	if (hierarchy) {
488 		path_size = strlen(node) + 10;
489 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
490 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
491 	} else {
492 		path_size = strlen(node) + 1;
493 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
494 		snprintf(path, path_size, "%s", node);
495 	}
496 
497 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
498 	if (err != 0) {
499 		message = "writing tx ring-ref";
500 		goto error;
501 	}
502 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
503 	if (err != 0) {
504 		message = "writing rx ring-ref";
505 		goto error;
506 	}
507 	err = xs_printf(*xst, path, "event-channel", "%u",
508 	    xen_intr_port(rxq->xen_intr_handle));
509 	if (err != 0) {
510 		message = "writing event-channel";
511 		goto error;
512 	}
513 
514 	free(path, M_DEVBUF);
515 
516 	return (0);
517 
518 error:
519 	free(path, M_DEVBUF);
520 	xenbus_dev_fatal(dev, err, "%s", message);
521 
522 	return (err);
523 }
524 
525 /* Common code used when first setting up, and when resuming. */
526 static int
527 talk_to_backend(device_t dev, struct netfront_info *info)
528 {
529 	const char *message;
530 	struct xs_transaction xst;
531 	const char *node = xenbus_get_node(dev);
532 	int err;
533 	unsigned long num_queues, max_queues = 0;
534 	unsigned int i;
535 
536 	err = xen_net_read_mac(dev, info->mac);
537 	if (err != 0) {
538 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
539 		goto out;
540 	}
541 
542 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
543 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
544 	if (err != 0)
545 		max_queues = 1;
546 	num_queues = xn_num_queues;
547 	if (num_queues > max_queues)
548 		num_queues = max_queues;
549 
550 	err = setup_device(dev, info, num_queues);
551 	if (err != 0)
552 		goto out;
553 
554  again:
555 	err = xs_transaction_start(&xst);
556 	if (err != 0) {
557 		xenbus_dev_fatal(dev, err, "starting transaction");
558 		goto free;
559 	}
560 
561 	if (info->num_queues == 1) {
562 		err = write_queue_xenstore_keys(dev, &info->rxq[0],
563 		    &info->txq[0], &xst, false);
564 		if (err != 0)
565 			goto abort_transaction_no_def_error;
566 	} else {
567 		err = xs_printf(xst, node, "multi-queue-num-queues",
568 		    "%u", info->num_queues);
569 		if (err != 0) {
570 			message = "writing multi-queue-num-queues";
571 			goto abort_transaction;
572 		}
573 
574 		for (i = 0; i < info->num_queues; i++) {
575 			err = write_queue_xenstore_keys(dev, &info->rxq[i],
576 			    &info->txq[i], &xst, true);
577 			if (err != 0)
578 				goto abort_transaction_no_def_error;
579 		}
580 	}
581 
582 	err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
583 	if (err != 0) {
584 		message = "writing request-rx-copy";
585 		goto abort_transaction;
586 	}
587 	err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
588 	if (err != 0) {
589 		message = "writing feature-rx-notify";
590 		goto abort_transaction;
591 	}
592 	err = xs_printf(xst, node, "feature-sg", "%d", 1);
593 	if (err != 0) {
594 		message = "writing feature-sg";
595 		goto abort_transaction;
596 	}
597 	err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
598 	if (err != 0) {
599 		message = "writing feature-gso-tcpv4";
600 		goto abort_transaction;
601 	}
602 
603 	err = xs_transaction_end(xst, 0);
604 	if (err != 0) {
605 		if (err == EAGAIN)
606 			goto again;
607 		xenbus_dev_fatal(dev, err, "completing transaction");
608 		goto free;
609 	}
610 
611 	return 0;
612 
613  abort_transaction:
614 	xenbus_dev_fatal(dev, err, "%s", message);
615  abort_transaction_no_def_error:
616 	xs_transaction_end(xst, 1);
617  free:
618 	netif_free(info);
619  out:
620 	return (err);
621 }
622 
623 static void
624 xn_rxq_tq_intr(void *xrxq, int pending)
625 {
626 	struct netfront_rxq *rxq = xrxq;
627 
628 	XN_RX_LOCK(rxq);
629 	xn_rxeof(rxq);
630 	XN_RX_UNLOCK(rxq);
631 }
632 
633 static void
634 xn_txq_start(struct netfront_txq *txq)
635 {
636 	struct netfront_info *np = txq->info;
637 	struct ifnet *ifp = np->xn_ifp;
638 
639 	XN_TX_LOCK_ASSERT(txq);
640 	if (!drbr_empty(ifp, txq->br))
641 		xn_txq_mq_start_locked(txq, NULL);
642 }
643 
644 static void
645 xn_txq_tq_intr(void *xtxq, int pending)
646 {
647 	struct netfront_txq *txq = xtxq;
648 
649 	XN_TX_LOCK(txq);
650 	if (RING_HAS_UNCONSUMED_RESPONSES(&txq->ring))
651 		xn_txeof(txq);
652 	xn_txq_start(txq);
653 	XN_TX_UNLOCK(txq);
654 }
655 
656 static void
657 xn_txq_tq_deferred(void *xtxq, int pending)
658 {
659 	struct netfront_txq *txq = xtxq;
660 
661 	XN_TX_LOCK(txq);
662 	xn_txq_start(txq);
663 	XN_TX_UNLOCK(txq);
664 }
665 
666 static void
667 disconnect_rxq(struct netfront_rxq *rxq)
668 {
669 
670 	xn_release_rx_bufs(rxq);
671 	gnttab_free_grant_references(rxq->gref_head);
672 	gnttab_end_foreign_access_ref(rxq->ring_ref);
673 	/*
674 	 * No split event channel support at the moment, handle will
675 	 * be unbound in tx. So no need to call xen_intr_unbind here,
676 	 * but we do want to reset the handler to 0.
677 	 */
678 	rxq->xen_intr_handle = 0;
679 }
680 
681 static void
682 destroy_rxq(struct netfront_rxq *rxq)
683 {
684 
685 	callout_drain(&rxq->rx_refill);
686 	free(rxq->ring.sring, M_DEVBUF);
687 	taskqueue_drain_all(rxq->tq);
688 	taskqueue_free(rxq->tq);
689 }
690 
691 static void
692 destroy_rxqs(struct netfront_info *np)
693 {
694 	int i;
695 
696 	for (i = 0; i < np->num_queues; i++)
697 		destroy_rxq(&np->rxq[i]);
698 
699 	free(np->rxq, M_DEVBUF);
700 	np->rxq = NULL;
701 }
702 
703 static int
704 setup_rxqs(device_t dev, struct netfront_info *info,
705 	   unsigned long num_queues)
706 {
707 	int q, i;
708 	int error;
709 	netif_rx_sring_t *rxs;
710 	struct netfront_rxq *rxq;
711 
712 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
713 	    M_DEVBUF, M_WAITOK|M_ZERO);
714 
715 	for (q = 0; q < num_queues; q++) {
716 		rxq = &info->rxq[q];
717 
718 		rxq->id = q;
719 		rxq->info = info;
720 		rxq->ring_ref = GRANT_REF_INVALID;
721 		rxq->ring.sring = NULL;
722 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
723 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
724 		    MTX_DEF);
725 
726 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
727 			rxq->mbufs[i] = NULL;
728 			rxq->grant_ref[i] = GRANT_REF_INVALID;
729 		}
730 
731 		/* Start resources allocation */
732 
733 		if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
734 		    &rxq->gref_head) != 0) {
735 			device_printf(dev, "allocating rx gref");
736 			error = ENOMEM;
737 			goto fail;
738 		}
739 
740 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
741 		    M_WAITOK|M_ZERO);
742 		SHARED_RING_INIT(rxs);
743 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
744 
745 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
746 		    &rxq->ring_ref);
747 		if (error != 0) {
748 			device_printf(dev, "granting rx ring page");
749 			goto fail_grant_ring;
750 		}
751 
752 		TASK_INIT(&rxq->intrtask, 0, xn_rxq_tq_intr, rxq);
753 		rxq->tq = taskqueue_create_fast(rxq->name, M_WAITOK,
754 		    taskqueue_thread_enqueue, &rxq->tq);
755 
756 		callout_init(&rxq->rx_refill, 1);
757 
758 		error = taskqueue_start_threads(&rxq->tq, 1, PI_NET,
759 		    "%s rxq %d", device_get_nameunit(dev), rxq->id);
760 		if (error != 0) {
761 			device_printf(dev, "failed to start rx taskq %d\n",
762 			    rxq->id);
763 			goto fail_start_thread;
764 		}
765 	}
766 
767 	return (0);
768 
769 fail_start_thread:
770 	gnttab_end_foreign_access_ref(rxq->ring_ref);
771 	taskqueue_drain_all(rxq->tq);
772 	taskqueue_free(rxq->tq);
773 fail_grant_ring:
774 	gnttab_free_grant_references(rxq->gref_head);
775 	free(rxq->ring.sring, M_DEVBUF);
776 fail:
777 	for (; q >= 0; q--) {
778 		disconnect_rxq(&info->rxq[q]);
779 		destroy_rxq(&info->rxq[q]);
780 	}
781 
782 	free(info->rxq, M_DEVBUF);
783 	return (error);
784 }
785 
786 static void
787 disconnect_txq(struct netfront_txq *txq)
788 {
789 
790 	xn_release_tx_bufs(txq);
791 	gnttab_free_grant_references(txq->gref_head);
792 	gnttab_end_foreign_access_ref(txq->ring_ref);
793 	xen_intr_unbind(&txq->xen_intr_handle);
794 }
795 
796 static void
797 destroy_txq(struct netfront_txq *txq)
798 {
799 
800 	free(txq->ring.sring, M_DEVBUF);
801 	buf_ring_free(txq->br, M_DEVBUF);
802 	taskqueue_drain_all(txq->tq);
803 	taskqueue_free(txq->tq);
804 }
805 
806 static void
807 destroy_txqs(struct netfront_info *np)
808 {
809 	int i;
810 
811 	for (i = 0; i < np->num_queues; i++)
812 		destroy_txq(&np->txq[i]);
813 
814 	free(np->txq, M_DEVBUF);
815 	np->txq = NULL;
816 }
817 
818 static int
819 setup_txqs(device_t dev, struct netfront_info *info,
820 	   unsigned long num_queues)
821 {
822 	int q, i;
823 	int error;
824 	netif_tx_sring_t *txs;
825 	struct netfront_txq *txq;
826 
827 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
828 	    M_DEVBUF, M_WAITOK|M_ZERO);
829 
830 	for (q = 0; q < num_queues; q++) {
831 		txq = &info->txq[q];
832 
833 		txq->id = q;
834 		txq->info = info;
835 
836 		txq->ring_ref = GRANT_REF_INVALID;
837 		txq->ring.sring = NULL;
838 
839 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
840 
841 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
842 		    MTX_DEF);
843 
844 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
845 			txq->mbufs[i] = (void *) ((u_long) i+1);
846 			txq->grant_ref[i] = GRANT_REF_INVALID;
847 		}
848 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
849 
850 		/* Start resources allocation. */
851 
852 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
853 		    &txq->gref_head) != 0) {
854 			device_printf(dev, "failed to allocate tx grant refs\n");
855 			error = ENOMEM;
856 			goto fail;
857 		}
858 
859 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
860 		    M_WAITOK|M_ZERO);
861 		SHARED_RING_INIT(txs);
862 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
863 
864 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
865 		    &txq->ring_ref);
866 		if (error != 0) {
867 			device_printf(dev, "failed to grant tx ring\n");
868 			goto fail_grant_ring;
869 		}
870 
871 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
872 		    M_WAITOK, &txq->lock);
873 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
874 		TASK_INIT(&txq->intrtask, 0, xn_txq_tq_intr, txq);
875 
876 		txq->tq = taskqueue_create_fast(txq->name, M_WAITOK,
877 		    taskqueue_thread_enqueue, &txq->tq);
878 
879 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
880 		    "%s txq %d", device_get_nameunit(dev), txq->id);
881 		if (error != 0) {
882 			device_printf(dev, "failed to start tx taskq %d\n",
883 			    txq->id);
884 			goto fail_start_thread;
885 		}
886 
887 		error = xen_intr_alloc_and_bind_local_port(dev,
888 			    xenbus_get_otherend_id(dev), xn_intr, /* handler */ NULL,
889 			    &info->txq[q],
890 			    INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
891 			    &txq->xen_intr_handle);
892 
893 		if (error != 0) {
894 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
895 			goto fail_bind_port;
896 		}
897 	}
898 
899 	return (0);
900 
901 fail_bind_port:
902 	taskqueue_drain_all(txq->tq);
903 fail_start_thread:
904 	buf_ring_free(txq->br, M_DEVBUF);
905 	taskqueue_free(txq->tq);
906 	gnttab_end_foreign_access_ref(txq->ring_ref);
907 fail_grant_ring:
908 	gnttab_free_grant_references(txq->gref_head);
909 	free(txq->ring.sring, M_DEVBUF);
910 fail:
911 	for (; q >= 0; q--) {
912 		disconnect_txq(&info->txq[q]);
913 		destroy_txq(&info->txq[q]);
914 	}
915 
916 	free(info->txq, M_DEVBUF);
917 	return (error);
918 }
919 
920 static int
921 setup_device(device_t dev, struct netfront_info *info,
922     unsigned long num_queues)
923 {
924 	int error;
925 	int q;
926 
927 	if (info->txq)
928 		destroy_txqs(info);
929 
930 	if (info->rxq)
931 		destroy_rxqs(info);
932 
933 	info->num_queues = 0;
934 
935 	error = setup_rxqs(dev, info, num_queues);
936 	if (error != 0)
937 		goto out;
938 	error = setup_txqs(dev, info, num_queues);
939 	if (error != 0)
940 		goto out;
941 
942 	info->num_queues = num_queues;
943 
944 	/* No split event channel at the moment. */
945 	for (q = 0; q < num_queues; q++)
946 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
947 
948 	return (0);
949 
950 out:
951 	KASSERT(error != 0, ("Error path taken without providing an error code"));
952 	return (error);
953 }
954 
955 #ifdef INET
956 /**
957  * If this interface has an ipv4 address, send an arp for it. This
958  * helps to get the network going again after migrating hosts.
959  */
960 static void
961 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
962 {
963 	struct ifnet *ifp;
964 	struct ifaddr *ifa;
965 
966 	ifp = info->xn_ifp;
967 	TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
968 		if (ifa->ifa_addr->sa_family == AF_INET) {
969 			arp_ifinit(ifp, ifa);
970 		}
971 	}
972 }
973 #endif
974 
975 /**
976  * Callback received when the backend's state changes.
977  */
978 static void
979 netfront_backend_changed(device_t dev, XenbusState newstate)
980 {
981 	struct netfront_info *sc = device_get_softc(dev);
982 
983 	DPRINTK("newstate=%d\n", newstate);
984 
985 	switch (newstate) {
986 	case XenbusStateInitialising:
987 	case XenbusStateInitialised:
988 	case XenbusStateUnknown:
989 	case XenbusStateClosed:
990 	case XenbusStateReconfigured:
991 	case XenbusStateReconfiguring:
992 		break;
993 	case XenbusStateInitWait:
994 		if (xenbus_get_state(dev) != XenbusStateInitialising)
995 			break;
996 		if (xn_connect(sc) != 0)
997 			break;
998 		xenbus_set_state(dev, XenbusStateConnected);
999 		break;
1000 	case XenbusStateClosing:
1001 		xenbus_set_state(dev, XenbusStateClosed);
1002 		break;
1003 	case XenbusStateConnected:
1004 #ifdef INET
1005 		netfront_send_fake_arp(dev, sc);
1006 #endif
1007 		break;
1008 	}
1009 }
1010 
1011 /**
1012  * \brief Verify that there is sufficient space in the Tx ring
1013  *        buffer for a maximally sized request to be enqueued.
1014  *
1015  * A transmit request requires a transmit descriptor for each packet
1016  * fragment, plus up to 2 entries for "options" (e.g. TSO).
1017  */
1018 static inline int
1019 xn_tx_slot_available(struct netfront_txq *txq)
1020 {
1021 
1022 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1023 }
1024 
1025 static void
1026 xn_release_tx_bufs(struct netfront_txq *txq)
1027 {
1028 	int i;
1029 
1030 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1031 		struct mbuf *m;
1032 
1033 		m = txq->mbufs[i];
1034 
1035 		/*
1036 		 * We assume that no kernel addresses are
1037 		 * less than NET_TX_RING_SIZE.  Any entry
1038 		 * in the table that is below this number
1039 		 * must be an index from free-list tracking.
1040 		 */
1041 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1042 			continue;
1043 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1044 		gnttab_release_grant_reference(&txq->gref_head,
1045 		    txq->grant_ref[i]);
1046 		txq->grant_ref[i] = GRANT_REF_INVALID;
1047 		add_id_to_freelist(txq->mbufs, i);
1048 		txq->mbufs_cnt--;
1049 		if (txq->mbufs_cnt < 0) {
1050 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1051 		}
1052 		m_free(m);
1053 	}
1054 }
1055 
1056 static struct mbuf *
1057 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1058 {
1059 	struct mbuf *m;
1060 
1061 	m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1062 	if (m == NULL)
1063 		return NULL;
1064 	m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1065 
1066 	return (m);
1067 }
1068 
1069 static void
1070 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1071 {
1072 	RING_IDX req_prod;
1073 	int notify;
1074 
1075 	XN_RX_LOCK_ASSERT(rxq);
1076 
1077 	if (__predict_false(rxq->info->carrier == 0))
1078 		return;
1079 
1080 	for (req_prod = rxq->ring.req_prod_pvt;
1081 	     req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1082 	     req_prod++) {
1083 		struct mbuf *m;
1084 		unsigned short id;
1085 		grant_ref_t ref;
1086 		struct netif_rx_request *req;
1087 		unsigned long pfn;
1088 
1089 		m = xn_alloc_one_rx_buffer(rxq);
1090 		if (m == NULL)
1091 			break;
1092 
1093 		id = xn_rxidx(req_prod);
1094 
1095 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1096 		rxq->mbufs[id] = m;
1097 
1098 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1099 		KASSERT(ref != GNTTAB_LIST_END,
1100 		    ("reserved grant references exhuasted"));
1101 		rxq->grant_ref[id] = ref;
1102 
1103 		pfn = atop(vtophys(mtod(m, vm_offset_t)));
1104 		req = RING_GET_REQUEST(&rxq->ring, req_prod);
1105 
1106 		gnttab_grant_foreign_access_ref(ref,
1107 		    xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1108 		req->id = id;
1109 		req->gref = ref;
1110 	}
1111 
1112 	rxq->ring.req_prod_pvt = req_prod;
1113 
1114 	/* Not enough requests? Try again later. */
1115 	if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1116 		callout_reset(&rxq->rx_refill, hz/10, xn_alloc_rx_buffers_callout,
1117 		    rxq);
1118 		return;
1119 	}
1120 
1121 	wmb();		/* barrier so backend seens requests */
1122 
1123 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1124 	if (notify)
1125 		xen_intr_signal(rxq->xen_intr_handle);
1126 }
1127 
1128 static void xn_alloc_rx_buffers_callout(void *arg)
1129 {
1130 	struct netfront_rxq *rxq;
1131 
1132 	rxq = (struct netfront_rxq *)arg;
1133 	XN_RX_LOCK(rxq);
1134 	xn_alloc_rx_buffers(rxq);
1135 	XN_RX_UNLOCK(rxq);
1136 }
1137 
1138 static void
1139 xn_release_rx_bufs(struct netfront_rxq *rxq)
1140 {
1141 	int i,  ref;
1142 	struct mbuf *m;
1143 
1144 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1145 		m = rxq->mbufs[i];
1146 
1147 		if (m == NULL)
1148 			continue;
1149 
1150 		ref = rxq->grant_ref[i];
1151 		if (ref == GRANT_REF_INVALID)
1152 			continue;
1153 
1154 		gnttab_end_foreign_access_ref(ref);
1155 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1156 		rxq->mbufs[i] = NULL;
1157 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1158 		m_freem(m);
1159 	}
1160 }
1161 
1162 static void
1163 xn_rxeof(struct netfront_rxq *rxq)
1164 {
1165 	struct ifnet *ifp;
1166 	struct netfront_info *np = rxq->info;
1167 #if (defined(INET) || defined(INET6))
1168 	struct lro_ctrl *lro = &rxq->lro;
1169 #endif
1170 	struct netfront_rx_info rinfo;
1171 	struct netif_rx_response *rx = &rinfo.rx;
1172 	struct netif_extra_info *extras = rinfo.extras;
1173 	RING_IDX i, rp;
1174 	struct mbuf *m;
1175 	struct mbufq mbufq_rxq, mbufq_errq;
1176 	int err, work_to_do;
1177 
1178 	do {
1179 		XN_RX_LOCK_ASSERT(rxq);
1180 		if (!netfront_carrier_ok(np))
1181 			return;
1182 
1183 		/* XXX: there should be some sane limit. */
1184 		mbufq_init(&mbufq_errq, INT_MAX);
1185 		mbufq_init(&mbufq_rxq, INT_MAX);
1186 
1187 		ifp = np->xn_ifp;
1188 
1189 		rp = rxq->ring.sring->rsp_prod;
1190 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1191 
1192 		i = rxq->ring.rsp_cons;
1193 		while ((i != rp)) {
1194 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1195 			memset(extras, 0, sizeof(rinfo.extras));
1196 
1197 			m = NULL;
1198 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1199 
1200 			if (__predict_false(err)) {
1201 				if (m)
1202 					(void )mbufq_enqueue(&mbufq_errq, m);
1203 				rxq->stats.rx_errors++;
1204 				continue;
1205 			}
1206 
1207 			m->m_pkthdr.rcvif = ifp;
1208 			if ( rx->flags & NETRXF_data_validated ) {
1209 				/* Tell the stack the checksums are okay */
1210 				/*
1211 				 * XXX this isn't necessarily the case - need to add
1212 				 * check
1213 				 */
1214 
1215 				m->m_pkthdr.csum_flags |=
1216 					(CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID
1217 					    | CSUM_PSEUDO_HDR);
1218 				m->m_pkthdr.csum_data = 0xffff;
1219 			}
1220 			if ((rx->flags & NETRXF_extra_info) != 0 &&
1221 			    (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1222 			    XEN_NETIF_EXTRA_TYPE_GSO)) {
1223 				m->m_pkthdr.tso_segsz =
1224 				extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1225 				m->m_pkthdr.csum_flags |= CSUM_TSO;
1226 			}
1227 
1228 			rxq->stats.rx_packets++;
1229 			rxq->stats.rx_bytes += m->m_pkthdr.len;
1230 
1231 			(void )mbufq_enqueue(&mbufq_rxq, m);
1232 			rxq->ring.rsp_cons = i;
1233 		}
1234 
1235 		mbufq_drain(&mbufq_errq);
1236 
1237 		/*
1238 		 * Process all the mbufs after the remapping is complete.
1239 		 * Break the mbuf chain first though.
1240 		 */
1241 		while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1242 			if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1243 
1244 			/* XXX: Do we really need to drop the rx lock? */
1245 			XN_RX_UNLOCK(rxq);
1246 #if (defined(INET) || defined(INET6))
1247 			/* Use LRO if possible */
1248 			if ((ifp->if_capenable & IFCAP_LRO) == 0 ||
1249 			    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1250 				/*
1251 				 * If LRO fails, pass up to the stack
1252 				 * directly.
1253 				 */
1254 				(*ifp->if_input)(ifp, m);
1255 			}
1256 #else
1257 			(*ifp->if_input)(ifp, m);
1258 #endif
1259 
1260 			XN_RX_LOCK(rxq);
1261 		}
1262 
1263 		rxq->ring.rsp_cons = i;
1264 
1265 #if (defined(INET) || defined(INET6))
1266 		/*
1267 		 * Flush any outstanding LRO work
1268 		 */
1269 		tcp_lro_flush_all(lro);
1270 #endif
1271 
1272 		xn_alloc_rx_buffers(rxq);
1273 
1274 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1275 	} while (work_to_do);
1276 }
1277 
1278 static void
1279 xn_txeof(struct netfront_txq *txq)
1280 {
1281 	RING_IDX i, prod;
1282 	unsigned short id;
1283 	struct ifnet *ifp;
1284 	netif_tx_response_t *txr;
1285 	struct mbuf *m;
1286 	struct netfront_info *np = txq->info;
1287 
1288 	XN_TX_LOCK_ASSERT(txq);
1289 
1290 	if (!netfront_carrier_ok(np))
1291 		return;
1292 
1293 	ifp = np->xn_ifp;
1294 
1295 	do {
1296 		prod = txq->ring.sring->rsp_prod;
1297 		rmb(); /* Ensure we see responses up to 'rp'. */
1298 
1299 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1300 			txr = RING_GET_RESPONSE(&txq->ring, i);
1301 			if (txr->status == NETIF_RSP_NULL)
1302 				continue;
1303 
1304 			if (txr->status != NETIF_RSP_OKAY) {
1305 				printf("%s: WARNING: response is %d!\n",
1306 				       __func__, txr->status);
1307 			}
1308 			id = txr->id;
1309 			m = txq->mbufs[id];
1310 			KASSERT(m != NULL, ("mbuf not found in chain"));
1311 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1312 				("mbuf already on the free list, but we're "
1313 				"trying to free it again!"));
1314 			M_ASSERTVALID(m);
1315 
1316 			/*
1317 			 * Increment packet count if this is the last
1318 			 * mbuf of the chain.
1319 			 */
1320 			if (!m->m_next)
1321 				if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1322 			if (__predict_false(gnttab_query_foreign_access(
1323 			    txq->grant_ref[id]) != 0)) {
1324 				panic("%s: grant id %u still in use by the "
1325 				    "backend", __func__, id);
1326 			}
1327 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1328 			gnttab_release_grant_reference(
1329 				&txq->gref_head, txq->grant_ref[id]);
1330 			txq->grant_ref[id] = GRANT_REF_INVALID;
1331 
1332 			txq->mbufs[id] = NULL;
1333 			add_id_to_freelist(txq->mbufs, id);
1334 			txq->mbufs_cnt--;
1335 			m_free(m);
1336 			/* Only mark the txq active if we've freed up at least one slot to try */
1337 			ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1338 		}
1339 		txq->ring.rsp_cons = prod;
1340 
1341 		/*
1342 		 * Set a new event, then check for race with update of
1343 		 * tx_cons. Note that it is essential to schedule a
1344 		 * callback, no matter how few buffers are pending. Even if
1345 		 * there is space in the transmit ring, higher layers may
1346 		 * be blocked because too much data is outstanding: in such
1347 		 * cases notification from Xen is likely to be the only kick
1348 		 * that we'll get.
1349 		 */
1350 		txq->ring.sring->rsp_event =
1351 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1352 
1353 		mb();
1354 	} while (prod != txq->ring.sring->rsp_prod);
1355 
1356 	if (txq->full &&
1357 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1358 		txq->full = false;
1359 		taskqueue_enqueue(txq->tq, &txq->intrtask);
1360 	}
1361 }
1362 
1363 
1364 static void
1365 xn_rxq_intr(void *xrxq)
1366 {
1367 	struct netfront_rxq *rxq = xrxq;
1368 
1369 	taskqueue_enqueue(rxq->tq, &rxq->intrtask);
1370 }
1371 
1372 static void
1373 xn_txq_intr(void *xtxq)
1374 {
1375 	struct netfront_txq *txq = xtxq;
1376 
1377 	taskqueue_enqueue(txq->tq, &txq->intrtask);
1378 }
1379 
1380 static int
1381 xn_intr(void *xsc)
1382 {
1383 	struct netfront_txq *txq = xsc;
1384 	struct netfront_info *np = txq->info;
1385 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1386 
1387 	/* kick both tx and rx */
1388 	xn_rxq_intr(rxq);
1389 	xn_txq_intr(txq);
1390 
1391 	return (FILTER_HANDLED);
1392 }
1393 
1394 static void
1395 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1396     grant_ref_t ref)
1397 {
1398 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1399 
1400 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1401 	rxq->mbufs[new] = m;
1402 	rxq->grant_ref[new] = ref;
1403 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1404 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1405 	rxq->ring.req_prod_pvt++;
1406 }
1407 
1408 static int
1409 xn_get_extras(struct netfront_rxq *rxq,
1410     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1411 {
1412 	struct netif_extra_info *extra;
1413 
1414 	int err = 0;
1415 
1416 	do {
1417 		struct mbuf *m;
1418 		grant_ref_t ref;
1419 
1420 		if (__predict_false(*cons + 1 == rp)) {
1421 			err = EINVAL;
1422 			break;
1423 		}
1424 
1425 		extra = (struct netif_extra_info *)
1426 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1427 
1428 		if (__predict_false(!extra->type ||
1429 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1430 			err = EINVAL;
1431 		} else {
1432 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1433 		}
1434 
1435 		m = xn_get_rx_mbuf(rxq, *cons);
1436 		ref = xn_get_rx_ref(rxq,  *cons);
1437 		xn_move_rx_slot(rxq, m, ref);
1438 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1439 
1440 	return err;
1441 }
1442 
1443 static int
1444 xn_get_responses(struct netfront_rxq *rxq,
1445     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1446     struct mbuf  **list)
1447 {
1448 	struct netif_rx_response *rx = &rinfo->rx;
1449 	struct netif_extra_info *extras = rinfo->extras;
1450 	struct mbuf *m, *m0, *m_prev;
1451 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1452 	RING_IDX ref_cons = *cons;
1453 	int frags = 1;
1454 	int err = 0;
1455 	u_long ret;
1456 
1457 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1458 
1459 	if (rx->flags & NETRXF_extra_info) {
1460 		err = xn_get_extras(rxq, extras, rp, cons);
1461 	}
1462 
1463 	if (m0 != NULL) {
1464 		m0->m_pkthdr.len = 0;
1465 		m0->m_next = NULL;
1466 	}
1467 
1468 	for (;;) {
1469 #if 0
1470 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1471 			rx->status, rx->offset, frags);
1472 #endif
1473 		if (__predict_false(rx->status < 0 ||
1474 			rx->offset + rx->status > PAGE_SIZE)) {
1475 
1476 			xn_move_rx_slot(rxq, m, ref);
1477 			if (m0 == m)
1478 				m0 = NULL;
1479 			m = NULL;
1480 			err = EINVAL;
1481 			goto next_skip_queue;
1482 		}
1483 
1484 		/*
1485 		 * This definitely indicates a bug, either in this driver or in
1486 		 * the backend driver. In future this should flag the bad
1487 		 * situation to the system controller to reboot the backed.
1488 		 */
1489 		if (ref == GRANT_REF_INVALID) {
1490 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1491 			err = EINVAL;
1492 			goto next;
1493 		}
1494 
1495 		ret = gnttab_end_foreign_access_ref(ref);
1496 		KASSERT(ret, ("Unable to end access to grant references"));
1497 
1498 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1499 
1500 next:
1501 		if (m == NULL)
1502 			break;
1503 
1504 		m->m_len = rx->status;
1505 		m->m_data += rx->offset;
1506 		m0->m_pkthdr.len += rx->status;
1507 
1508 next_skip_queue:
1509 		if (!(rx->flags & NETRXF_more_data))
1510 			break;
1511 
1512 		if (*cons + frags == rp) {
1513 			if (net_ratelimit())
1514 				WPRINTK("Need more frags\n");
1515 			err = ENOENT;
1516 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1517 			       __func__, *cons, frags, rp);
1518 			break;
1519 		}
1520 		/*
1521 		 * Note that m can be NULL, if rx->status < 0 or if
1522 		 * rx->offset + rx->status > PAGE_SIZE above.
1523 		 */
1524 		m_prev = m;
1525 
1526 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1527 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1528 
1529 		/*
1530 		 * m_prev == NULL can happen if rx->status < 0 or if
1531 		 * rx->offset + * rx->status > PAGE_SIZE above.
1532 		 */
1533 		if (m_prev != NULL)
1534 			m_prev->m_next = m;
1535 
1536 		/*
1537 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1538 		 * rx->status > PAGE_SIZE above.
1539 		 */
1540 		if (m0 == NULL)
1541 			m0 = m;
1542 		m->m_next = NULL;
1543 		ref = xn_get_rx_ref(rxq, *cons + frags);
1544 		ref_cons = *cons + frags;
1545 		frags++;
1546 	}
1547 	*list = m0;
1548 	*cons += frags;
1549 
1550 	return (err);
1551 }
1552 
1553 /**
1554  * \brief Count the number of fragments in an mbuf chain.
1555  *
1556  * Surprisingly, there isn't an M* macro for this.
1557  */
1558 static inline int
1559 xn_count_frags(struct mbuf *m)
1560 {
1561 	int nfrags;
1562 
1563 	for (nfrags = 0; m != NULL; m = m->m_next)
1564 		nfrags++;
1565 
1566 	return (nfrags);
1567 }
1568 
1569 /**
1570  * Given an mbuf chain, make sure we have enough room and then push
1571  * it onto the transmit ring.
1572  */
1573 static int
1574 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1575 {
1576 	struct mbuf *m;
1577 	struct netfront_info *np = txq->info;
1578 	struct ifnet *ifp = np->xn_ifp;
1579 	u_int nfrags;
1580 	int otherend_id;
1581 
1582 	/**
1583 	 * Defragment the mbuf if necessary.
1584 	 */
1585 	nfrags = xn_count_frags(m_head);
1586 
1587 	/*
1588 	 * Check to see whether this request is longer than netback
1589 	 * can handle, and try to defrag it.
1590 	 */
1591 	/**
1592 	 * It is a bit lame, but the netback driver in Linux can't
1593 	 * deal with nfrags > MAX_TX_REQ_FRAGS, which is a quirk of
1594 	 * the Linux network stack.
1595 	 */
1596 	if (nfrags > np->maxfrags) {
1597 		m = m_defrag(m_head, M_NOWAIT);
1598 		if (!m) {
1599 			/*
1600 			 * Defrag failed, so free the mbuf and
1601 			 * therefore drop the packet.
1602 			 */
1603 			m_freem(m_head);
1604 			return (EMSGSIZE);
1605 		}
1606 		m_head = m;
1607 	}
1608 
1609 	/* Determine how many fragments now exist */
1610 	nfrags = xn_count_frags(m_head);
1611 
1612 	/*
1613 	 * Check to see whether the defragmented packet has too many
1614 	 * segments for the Linux netback driver.
1615 	 */
1616 	/**
1617 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1618 	 * of mbufs longer than Linux can handle.  Make sure we don't
1619 	 * pass a too-long chain over to the other side by dropping the
1620 	 * packet.  It doesn't look like there is currently a way to
1621 	 * tell the TCP stack to generate a shorter chain of packets.
1622 	 */
1623 	if (nfrags > MAX_TX_REQ_FRAGS) {
1624 #ifdef DEBUG
1625 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1626 		       "won't be able to handle it, dropping\n",
1627 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1628 #endif
1629 		m_freem(m_head);
1630 		return (EMSGSIZE);
1631 	}
1632 
1633 	/*
1634 	 * This check should be redundant.  We've already verified that we
1635 	 * have enough slots in the ring to handle a packet of maximum
1636 	 * size, and that our packet is less than the maximum size.  Keep
1637 	 * it in here as an assert for now just to make certain that
1638 	 * chain_cnt is accurate.
1639 	 */
1640 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1641 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1642 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1643                     (int) nfrags, (int) NET_TX_RING_SIZE));
1644 
1645 	/*
1646 	 * Start packing the mbufs in this chain into
1647 	 * the fragment pointers. Stop when we run out
1648 	 * of fragments or hit the end of the mbuf chain.
1649 	 */
1650 	m = m_head;
1651 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1652 	for (m = m_head; m; m = m->m_next) {
1653 		netif_tx_request_t *tx;
1654 		uintptr_t id;
1655 		grant_ref_t ref;
1656 		u_long mfn; /* XXX Wrong type? */
1657 
1658 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1659 		id = get_id_from_freelist(txq->mbufs);
1660 		if (id == 0)
1661 			panic("%s: was allocated the freelist head!\n",
1662 			    __func__);
1663 		txq->mbufs_cnt++;
1664 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1665 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1666 			    __func__);
1667 		txq->mbufs[id] = m;
1668 		tx->id = id;
1669 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1670 		KASSERT((short)ref >= 0, ("Negative ref"));
1671 		mfn = virt_to_mfn(mtod(m, vm_offset_t));
1672 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1673 		    mfn, GNTMAP_readonly);
1674 		tx->gref = txq->grant_ref[id] = ref;
1675 		tx->offset = mtod(m, vm_offset_t) & (PAGE_SIZE - 1);
1676 		tx->flags = 0;
1677 		if (m == m_head) {
1678 			/*
1679 			 * The first fragment has the entire packet
1680 			 * size, subsequent fragments have just the
1681 			 * fragment size. The backend works out the
1682 			 * true size of the first fragment by
1683 			 * subtracting the sizes of the other
1684 			 * fragments.
1685 			 */
1686 			tx->size = m->m_pkthdr.len;
1687 
1688 			/*
1689 			 * The first fragment contains the checksum flags
1690 			 * and is optionally followed by extra data for
1691 			 * TSO etc.
1692 			 */
1693 			/**
1694 			 * CSUM_TSO requires checksum offloading.
1695 			 * Some versions of FreeBSD fail to
1696 			 * set CSUM_TCP in the CSUM_TSO case,
1697 			 * so we have to test for CSUM_TSO
1698 			 * explicitly.
1699 			 */
1700 			if (m->m_pkthdr.csum_flags
1701 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1702 				tx->flags |= (NETTXF_csum_blank
1703 				    | NETTXF_data_validated);
1704 			}
1705 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1706 				struct netif_extra_info *gso =
1707 					(struct netif_extra_info *)
1708 					RING_GET_REQUEST(&txq->ring,
1709 							 ++txq->ring.req_prod_pvt);
1710 
1711 				tx->flags |= NETTXF_extra_info;
1712 
1713 				gso->u.gso.size = m->m_pkthdr.tso_segsz;
1714 				gso->u.gso.type =
1715 					XEN_NETIF_GSO_TYPE_TCPV4;
1716 				gso->u.gso.pad = 0;
1717 				gso->u.gso.features = 0;
1718 
1719 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1720 				gso->flags = 0;
1721 			}
1722 		} else {
1723 			tx->size = m->m_len;
1724 		}
1725 		if (m->m_next)
1726 			tx->flags |= NETTXF_more_data;
1727 
1728 		txq->ring.req_prod_pvt++;
1729 	}
1730 	BPF_MTAP(ifp, m_head);
1731 
1732 	xn_txeof(txq);
1733 
1734 	txq->stats.tx_bytes += m_head->m_pkthdr.len;
1735 	txq->stats.tx_packets++;
1736 
1737 	return (0);
1738 }
1739 
1740 /* equivalent of network_open() in Linux */
1741 static void
1742 xn_ifinit_locked(struct netfront_info *np)
1743 {
1744 	struct ifnet *ifp;
1745 	int i;
1746 	struct netfront_rxq *rxq;
1747 
1748 	XN_LOCK_ASSERT(np);
1749 
1750 	ifp = np->xn_ifp;
1751 
1752 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1753 		return;
1754 
1755 	xn_stop(np);
1756 
1757 	for (i = 0; i < np->num_queues; i++) {
1758 		rxq = &np->rxq[i];
1759 		XN_RX_LOCK(rxq);
1760 		xn_alloc_rx_buffers(rxq);
1761 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1762 		if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1763 			taskqueue_enqueue(rxq->tq, &rxq->intrtask);
1764 		XN_RX_UNLOCK(rxq);
1765 	}
1766 
1767 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
1768 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1769 	if_link_state_change(ifp, LINK_STATE_UP);
1770 }
1771 
1772 static void
1773 xn_ifinit(void *xsc)
1774 {
1775 	struct netfront_info *sc = xsc;
1776 
1777 	XN_LOCK(sc);
1778 	xn_ifinit_locked(sc);
1779 	XN_UNLOCK(sc);
1780 }
1781 
1782 static int
1783 xn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
1784 {
1785 	struct netfront_info *sc = ifp->if_softc;
1786 	struct ifreq *ifr = (struct ifreq *) data;
1787 #ifdef INET
1788 	struct ifaddr *ifa = (struct ifaddr *)data;
1789 #endif
1790 
1791 	int mask, error = 0;
1792 	switch(cmd) {
1793 	case SIOCSIFADDR:
1794 #ifdef INET
1795 		XN_LOCK(sc);
1796 		if (ifa->ifa_addr->sa_family == AF_INET) {
1797 			ifp->if_flags |= IFF_UP;
1798 			if (!(ifp->if_drv_flags & IFF_DRV_RUNNING))
1799 				xn_ifinit_locked(sc);
1800 			arp_ifinit(ifp, ifa);
1801 			XN_UNLOCK(sc);
1802 		} else {
1803 			XN_UNLOCK(sc);
1804 #endif
1805 			error = ether_ioctl(ifp, cmd, data);
1806 #ifdef INET
1807 		}
1808 #endif
1809 		break;
1810 	case SIOCSIFMTU:
1811 		ifp->if_mtu = ifr->ifr_mtu;
1812 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1813 		xn_ifinit(sc);
1814 		break;
1815 	case SIOCSIFFLAGS:
1816 		XN_LOCK(sc);
1817 		if (ifp->if_flags & IFF_UP) {
1818 			/*
1819 			 * If only the state of the PROMISC flag changed,
1820 			 * then just use the 'set promisc mode' command
1821 			 * instead of reinitializing the entire NIC. Doing
1822 			 * a full re-init means reloading the firmware and
1823 			 * waiting for it to start up, which may take a
1824 			 * second or two.
1825 			 */
1826 			xn_ifinit_locked(sc);
1827 		} else {
1828 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1829 				xn_stop(sc);
1830 			}
1831 		}
1832 		sc->xn_if_flags = ifp->if_flags;
1833 		XN_UNLOCK(sc);
1834 		break;
1835 	case SIOCSIFCAP:
1836 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
1837 		if (mask & IFCAP_TXCSUM) {
1838 			if (IFCAP_TXCSUM & ifp->if_capenable) {
1839 				ifp->if_capenable &= ~(IFCAP_TXCSUM|IFCAP_TSO4);
1840 				ifp->if_hwassist &= ~(CSUM_TCP | CSUM_UDP
1841 				    | CSUM_IP | CSUM_TSO);
1842 			} else {
1843 				ifp->if_capenable |= IFCAP_TXCSUM;
1844 				ifp->if_hwassist |= (CSUM_TCP | CSUM_UDP
1845 				    | CSUM_IP);
1846 			}
1847 		}
1848 		if (mask & IFCAP_RXCSUM) {
1849 			ifp->if_capenable ^= IFCAP_RXCSUM;
1850 		}
1851 		if (mask & IFCAP_TSO4) {
1852 			if (IFCAP_TSO4 & ifp->if_capenable) {
1853 				ifp->if_capenable &= ~IFCAP_TSO4;
1854 				ifp->if_hwassist &= ~CSUM_TSO;
1855 			} else if (IFCAP_TXCSUM & ifp->if_capenable) {
1856 				ifp->if_capenable |= IFCAP_TSO4;
1857 				ifp->if_hwassist |= CSUM_TSO;
1858 			} else {
1859 				IPRINTK("Xen requires tx checksum offload"
1860 				    " be enabled to use TSO\n");
1861 				error = EINVAL;
1862 			}
1863 		}
1864 		if (mask & IFCAP_LRO) {
1865 			ifp->if_capenable ^= IFCAP_LRO;
1866 
1867 		}
1868 		break;
1869 	case SIOCADDMULTI:
1870 	case SIOCDELMULTI:
1871 		break;
1872 	case SIOCSIFMEDIA:
1873 	case SIOCGIFMEDIA:
1874 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1875 		break;
1876 	default:
1877 		error = ether_ioctl(ifp, cmd, data);
1878 	}
1879 
1880 	return (error);
1881 }
1882 
1883 static void
1884 xn_stop(struct netfront_info *sc)
1885 {
1886 	struct ifnet *ifp;
1887 
1888 	XN_LOCK_ASSERT(sc);
1889 
1890 	ifp = sc->xn_ifp;
1891 
1892 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1893 	if_link_state_change(ifp, LINK_STATE_DOWN);
1894 }
1895 
1896 static void
1897 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1898 {
1899 	int requeue_idx, i;
1900 	grant_ref_t ref;
1901 	netif_rx_request_t *req;
1902 
1903 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1904 		struct mbuf *m;
1905 		u_long pfn;
1906 
1907 		if (rxq->mbufs[i] == NULL)
1908 			continue;
1909 
1910 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1911 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1912 
1913 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1914 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1915 
1916 		gnttab_grant_foreign_access_ref(ref,
1917 		    xenbus_get_otherend_id(rxq->info->xbdev),
1918 		    pfn, 0);
1919 
1920 		req->gref = ref;
1921 		req->id   = requeue_idx;
1922 
1923 		requeue_idx++;
1924 	}
1925 
1926 	rxq->ring.req_prod_pvt = requeue_idx;
1927 }
1928 
1929 /* START of Xenolinux helper functions adapted to FreeBSD */
1930 int
1931 xn_connect(struct netfront_info *np)
1932 {
1933 	int i, error;
1934 	u_int feature_rx_copy;
1935 	struct netfront_rxq *rxq;
1936 	struct netfront_txq *txq;
1937 
1938 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1939 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
1940 	if (error != 0)
1941 		feature_rx_copy = 0;
1942 
1943 	/* We only support rx copy. */
1944 	if (!feature_rx_copy)
1945 		return (EPROTONOSUPPORT);
1946 
1947 	/* Recovery procedure: */
1948 	error = talk_to_backend(np->xbdev, np);
1949 	if (error != 0)
1950 		return (error);
1951 
1952 	/* Step 1: Reinitialise variables. */
1953 	xn_query_features(np);
1954 	xn_configure_features(np);
1955 
1956 	/* Step 2: Release TX buffer */
1957 	for (i = 0; i < np->num_queues; i++) {
1958 		txq = &np->txq[i];
1959 		xn_release_tx_bufs(txq);
1960 	}
1961 
1962 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
1963 	for (i = 0; i < np->num_queues; i++) {
1964 		rxq = &np->rxq[i];
1965 		xn_rebuild_rx_bufs(rxq);
1966 	}
1967 
1968 	/* Step 4: All public and private state should now be sane.  Get
1969 	 * ready to start sending and receiving packets and give the driver
1970 	 * domain a kick because we've probably just requeued some
1971 	 * packets.
1972 	 */
1973 	netfront_carrier_on(np);
1974 	for (i = 0; i < np->num_queues; i++) {
1975 		txq = &np->txq[i];
1976 		xen_intr_signal(txq->xen_intr_handle);
1977 		XN_TX_LOCK(txq);
1978 		xn_txeof(txq);
1979 		XN_TX_UNLOCK(txq);
1980 		XN_RX_LOCK(rxq);
1981 		xn_alloc_rx_buffers(rxq);
1982 		XN_RX_UNLOCK(rxq);
1983 	}
1984 
1985 	return (0);
1986 }
1987 
1988 static void
1989 xn_query_features(struct netfront_info *np)
1990 {
1991 	int val;
1992 
1993 	device_printf(np->xbdev, "backend features:");
1994 
1995 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1996 		"feature-sg", NULL, "%d", &val) != 0)
1997 		val = 0;
1998 
1999 	np->maxfrags = 1;
2000 	if (val) {
2001 		np->maxfrags = MAX_TX_REQ_FRAGS;
2002 		printf(" feature-sg");
2003 	}
2004 
2005 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2006 		"feature-gso-tcpv4", NULL, "%d", &val) != 0)
2007 		val = 0;
2008 
2009 	np->xn_ifp->if_capabilities &= ~(IFCAP_TSO4|IFCAP_LRO);
2010 	if (val) {
2011 		np->xn_ifp->if_capabilities |= IFCAP_TSO4|IFCAP_LRO;
2012 		printf(" feature-gso-tcp4");
2013 	}
2014 
2015 	printf("\n");
2016 }
2017 
2018 static int
2019 xn_configure_features(struct netfront_info *np)
2020 {
2021 	int err, cap_enabled;
2022 #if (defined(INET) || defined(INET6))
2023 	int i;
2024 #endif
2025 
2026 	err = 0;
2027 
2028 	if (np->xn_resume &&
2029 	    ((np->xn_ifp->if_capenable & np->xn_ifp->if_capabilities)
2030 	    == np->xn_ifp->if_capenable)) {
2031 		/* Current options are available, no need to do anything. */
2032 		return (0);
2033 	}
2034 
2035 	/* Try to preserve as many options as possible. */
2036 	if (np->xn_resume)
2037 		cap_enabled = np->xn_ifp->if_capenable;
2038 	else
2039 		cap_enabled = UINT_MAX;
2040 
2041 #if (defined(INET) || defined(INET6))
2042 	for (i = 0; i < np->num_queues; i++)
2043 		if ((np->xn_ifp->if_capenable & IFCAP_LRO) ==
2044 		    (cap_enabled & IFCAP_LRO))
2045 			tcp_lro_free(&np->rxq[i].lro);
2046 #endif
2047     	np->xn_ifp->if_capenable =
2048 	    np->xn_ifp->if_capabilities & ~(IFCAP_LRO|IFCAP_TSO4) & cap_enabled;
2049 	np->xn_ifp->if_hwassist &= ~CSUM_TSO;
2050 #if (defined(INET) || defined(INET6))
2051 	for (i = 0; i < np->num_queues; i++) {
2052 		if (xn_enable_lro && (np->xn_ifp->if_capabilities & IFCAP_LRO) ==
2053 		    (cap_enabled & IFCAP_LRO)) {
2054 			err = tcp_lro_init(&np->rxq[i].lro);
2055 			if (err != 0) {
2056 				device_printf(np->xbdev, "LRO initialization failed\n");
2057 			} else {
2058 				np->rxq[i].lro.ifp = np->xn_ifp;
2059 				np->xn_ifp->if_capenable |= IFCAP_LRO;
2060 			}
2061 		}
2062 	}
2063 	if ((np->xn_ifp->if_capabilities & IFCAP_TSO4) ==
2064 	    (cap_enabled & IFCAP_TSO4)) {
2065 		np->xn_ifp->if_capenable |= IFCAP_TSO4;
2066 		np->xn_ifp->if_hwassist |= CSUM_TSO;
2067 	}
2068 #endif
2069 	return (err);
2070 }
2071 
2072 static int
2073 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2074 {
2075 	struct netfront_info *np;
2076 	struct ifnet *ifp;
2077 	struct buf_ring *br;
2078 	int error, notify;
2079 
2080 	np = txq->info;
2081 	br = txq->br;
2082 	ifp = np->xn_ifp;
2083 	error = 0;
2084 
2085 	XN_TX_LOCK_ASSERT(txq);
2086 
2087 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
2088 	    !netfront_carrier_ok(np)) {
2089 		if (m != NULL)
2090 			error = drbr_enqueue(ifp, br, m);
2091 		return (error);
2092 	}
2093 
2094 	if (m != NULL) {
2095 		error = drbr_enqueue(ifp, br, m);
2096 		if (error != 0)
2097 			return (error);
2098 	}
2099 
2100 	while ((m = drbr_peek(ifp, br)) != NULL) {
2101 		if (!xn_tx_slot_available(txq)) {
2102 			drbr_putback(ifp, br, m);
2103 			break;
2104 		}
2105 
2106 		error = xn_assemble_tx_request(txq, m);
2107 		/* xn_assemble_tx_request always consumes the mbuf*/
2108 		if (error != 0) {
2109 			drbr_advance(ifp, br);
2110 			break;
2111 		}
2112 
2113 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2114 		if (notify)
2115 			xen_intr_signal(txq->xen_intr_handle);
2116 
2117 		drbr_advance(ifp, br);
2118 	}
2119 
2120 	if (RING_FULL(&txq->ring))
2121 		txq->full = true;
2122 
2123 	return (0);
2124 }
2125 
2126 static int
2127 xn_txq_mq_start(struct ifnet *ifp, struct mbuf *m)
2128 {
2129 	struct netfront_info *np;
2130 	struct netfront_txq *txq;
2131 	int i, npairs, error;
2132 
2133 	np = ifp->if_softc;
2134 	npairs = np->num_queues;
2135 
2136 	/* check if flowid is set */
2137 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2138 		i = m->m_pkthdr.flowid % npairs;
2139 	else
2140 		i = curcpu % npairs;
2141 
2142 	txq = &np->txq[i];
2143 
2144 	if (XN_TX_TRYLOCK(txq) != 0) {
2145 		error = xn_txq_mq_start_locked(txq, m);
2146 		XN_TX_UNLOCK(txq);
2147 	} else {
2148 		error = drbr_enqueue(ifp, txq->br, m);
2149 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2150 	}
2151 
2152 	return (error);
2153 }
2154 
2155 static void
2156 xn_qflush(struct ifnet *ifp)
2157 {
2158 	struct netfront_info *np;
2159 	struct netfront_txq *txq;
2160 	struct mbuf *m;
2161 	int i;
2162 
2163 	np = ifp->if_softc;
2164 
2165 	for (i = 0; i < np->num_queues; i++) {
2166 		txq = &np->txq[i];
2167 
2168 		XN_TX_LOCK(txq);
2169 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2170 			m_freem(m);
2171 		XN_TX_UNLOCK(txq);
2172 	}
2173 
2174 	if_qflush(ifp);
2175 }
2176 
2177 /**
2178  * Create a network device.
2179  * @param dev  Newbus device representing this virtual NIC.
2180  */
2181 int
2182 create_netdev(device_t dev)
2183 {
2184 	struct netfront_info *np;
2185 	int err;
2186 	struct ifnet *ifp;
2187 
2188 	np = device_get_softc(dev);
2189 
2190 	np->xbdev         = dev;
2191 
2192 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2193 
2194 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2195 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2196 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2197 
2198 	err = xen_net_read_mac(dev, np->mac);
2199 	if (err != 0)
2200 		goto error;
2201 
2202 	/* Set up ifnet structure */
2203 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2204     	ifp->if_softc = np;
2205     	if_initname(ifp, "xn",  device_get_unit(dev));
2206     	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2207     	ifp->if_ioctl = xn_ioctl;
2208 
2209 	ifp->if_transmit = xn_txq_mq_start;
2210 	ifp->if_qflush = xn_qflush;
2211 
2212     	ifp->if_init = xn_ifinit;
2213 
2214     	ifp->if_hwassist = XN_CSUM_FEATURES;
2215     	ifp->if_capabilities = IFCAP_HWCSUM;
2216 	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2217 	ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
2218 	ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
2219 
2220     	ether_ifattach(ifp, np->mac);
2221 	netfront_carrier_off(np);
2222 
2223 	return (0);
2224 
2225 error:
2226 	KASSERT(err != 0, ("Error path with no error code specified"));
2227 	return (err);
2228 }
2229 
2230 static int
2231 netfront_detach(device_t dev)
2232 {
2233 	struct netfront_info *info = device_get_softc(dev);
2234 
2235 	DPRINTK("%s\n", xenbus_get_node(dev));
2236 
2237 	netif_free(info);
2238 
2239 	return 0;
2240 }
2241 
2242 static void
2243 netif_free(struct netfront_info *np)
2244 {
2245 
2246 	XN_LOCK(np);
2247 	xn_stop(np);
2248 	XN_UNLOCK(np);
2249 	netif_disconnect_backend(np);
2250 	free(np->rxq, M_DEVBUF);
2251 	free(np->txq, M_DEVBUF);
2252 	ether_ifdetach(np->xn_ifp);
2253 	if_free(np->xn_ifp);
2254 	np->xn_ifp = NULL;
2255 	ifmedia_removeall(&np->sc_media);
2256 }
2257 
2258 static void
2259 netif_disconnect_backend(struct netfront_info *np)
2260 {
2261 	u_int i;
2262 
2263 	for (i = 0; i < np->num_queues; i++) {
2264 		XN_RX_LOCK(&np->rxq[i]);
2265 		XN_TX_LOCK(&np->txq[i]);
2266 	}
2267 	netfront_carrier_off(np);
2268 	for (i = 0; i < np->num_queues; i++) {
2269 		XN_RX_UNLOCK(&np->rxq[i]);
2270 		XN_TX_UNLOCK(&np->txq[i]);
2271 	}
2272 
2273 	for (i = 0; i < np->num_queues; i++) {
2274 		disconnect_rxq(&np->rxq[i]);
2275 		disconnect_txq(&np->txq[i]);
2276 	}
2277 }
2278 
2279 static int
2280 xn_ifmedia_upd(struct ifnet *ifp)
2281 {
2282 
2283 	return (0);
2284 }
2285 
2286 static void
2287 xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
2288 {
2289 
2290 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2291 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2292 }
2293 
2294 /* ** Driver registration ** */
2295 static device_method_t netfront_methods[] = {
2296 	/* Device interface */
2297 	DEVMETHOD(device_probe,         netfront_probe),
2298 	DEVMETHOD(device_attach,        netfront_attach),
2299 	DEVMETHOD(device_detach,        netfront_detach),
2300 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2301 	DEVMETHOD(device_suspend,       netfront_suspend),
2302 	DEVMETHOD(device_resume,        netfront_resume),
2303 
2304 	/* Xenbus interface */
2305 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2306 
2307 	DEVMETHOD_END
2308 };
2309 
2310 static driver_t netfront_driver = {
2311 	"xn",
2312 	netfront_methods,
2313 	sizeof(struct netfront_info),
2314 };
2315 devclass_t netfront_devclass;
2316 
2317 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, netfront_devclass, NULL,
2318     NULL);
2319