xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
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 /*
81  * Should the driver do LRO on the RX end
82  *  this can be toggled on the fly, but the
83  *  interface must be reset (down/up) for it
84  *  to take effect.
85  */
86 static int xn_enable_lro = 1;
87 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
88 
89 /*
90  * Number of pairs of queues.
91  */
92 static unsigned long xn_num_queues = 4;
93 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
94 
95 /**
96  * \brief The maximum allowed data fragments in a single transmit
97  *        request.
98  *
99  * This limit is imposed by the backend driver.  We assume here that
100  * we are dealing with a Linux driver domain and have set our limit
101  * to mirror the Linux MAX_SKB_FRAGS constant.
102  */
103 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
104 
105 #define RX_COPY_THRESHOLD 256
106 
107 #define net_ratelimit() 0
108 
109 struct netfront_rxq;
110 struct netfront_txq;
111 struct netfront_info;
112 struct netfront_rx_info;
113 
114 static void xn_txeof(struct netfront_txq *);
115 static void xn_rxeof(struct netfront_rxq *);
116 static void xn_alloc_rx_buffers(struct netfront_rxq *);
117 
118 static void xn_release_rx_bufs(struct netfront_rxq *);
119 static void xn_release_tx_bufs(struct netfront_txq *);
120 
121 static void xn_rxq_intr(void *);
122 static void xn_txq_intr(void *);
123 static int  xn_intr(void *);
124 static inline int xn_count_frags(struct mbuf *m);
125 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
126 static int xn_ioctl(struct ifnet *, u_long, caddr_t);
127 static void xn_ifinit_locked(struct netfront_info *);
128 static void xn_ifinit(void *);
129 static void xn_stop(struct netfront_info *);
130 static void xn_query_features(struct netfront_info *np);
131 static int xn_configure_features(struct netfront_info *np);
132 static void netif_free(struct netfront_info *info);
133 static int netfront_detach(device_t dev);
134 
135 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
136 static int xn_txq_mq_start(struct ifnet *, struct mbuf *);
137 
138 static int talk_to_backend(device_t dev, struct netfront_info *info);
139 static int create_netdev(device_t dev);
140 static void netif_disconnect_backend(struct netfront_info *info);
141 static int setup_device(device_t dev, struct netfront_info *info,
142     unsigned long);
143 static int xn_ifmedia_upd(struct ifnet *ifp);
144 static void xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr);
145 
146 int xn_connect(struct netfront_info *);
147 
148 static int xn_get_responses(struct netfront_rxq *,
149     struct netfront_rx_info *, RING_IDX, RING_IDX *,
150     struct mbuf **);
151 
152 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
153 
154 #define INVALID_P2M_ENTRY (~0UL)
155 
156 struct xn_rx_stats
157 {
158 	u_long	rx_packets;	/* total packets received	*/
159 	u_long	rx_bytes;	/* total bytes received 	*/
160 	u_long	rx_errors;	/* bad packets received		*/
161 };
162 
163 struct xn_tx_stats
164 {
165 	u_long	tx_packets;	/* total packets transmitted	*/
166 	u_long	tx_bytes;	/* total bytes transmitted	*/
167 	u_long	tx_errors;	/* packet transmit problems	*/
168 };
169 
170 #define XN_QUEUE_NAME_LEN  8	/* xn{t,r}x_%u, allow for two digits */
171 struct netfront_rxq {
172 	struct netfront_info 	*info;
173 	u_int			id;
174 	char			name[XN_QUEUE_NAME_LEN];
175 	struct mtx		lock;
176 
177 	int			ring_ref;
178 	netif_rx_front_ring_t 	ring;
179 	xen_intr_handle_t	xen_intr_handle;
180 
181 	grant_ref_t 		gref_head;
182 	grant_ref_t 		grant_ref[NET_TX_RING_SIZE + 1];
183 
184 	struct mbuf		*mbufs[NET_RX_RING_SIZE + 1];
185 	struct mbufq		batch;		/* batch queue */
186 	int                     target;
187 
188 	xen_pfn_t		pfn_array[NET_RX_RING_SIZE];
189 
190 	struct lro_ctrl		lro;
191 
192 	struct taskqueue 	*tq;
193 	struct task		intrtask;
194 
195 	struct xn_rx_stats	stats;
196 };
197 
198 struct netfront_txq {
199 	struct netfront_info 	*info;
200 	u_int 			id;
201 	char			name[XN_QUEUE_NAME_LEN];
202 	struct mtx		lock;
203 
204 	int			ring_ref;
205 	netif_tx_front_ring_t	ring;
206 	xen_intr_handle_t 	xen_intr_handle;
207 
208 	grant_ref_t		gref_head;
209 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
210 
211 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
212 	int			mbufs_cnt;
213 	struct buf_ring		*br;
214 
215 	struct taskqueue 	*tq;
216 	struct task       	intrtask;
217 	struct task       	defrtask;
218 
219 	bool			full;
220 
221 	struct xn_tx_stats	stats;
222 };
223 
224 struct netfront_info {
225 	struct ifnet 		*xn_ifp;
226 
227 	struct mtx   		sc_lock;
228 
229 	u_int  num_queues;
230 	struct netfront_rxq 	*rxq;
231 	struct netfront_txq 	*txq;
232 
233 	u_int			carrier;
234 	u_int			maxfrags;
235 
236 	/* Receive-ring batched refills. */
237 #define RX_MIN_TARGET 32
238 #define RX_MAX_TARGET NET_RX_RING_SIZE
239 	int			rx_min_target;
240 	int			rx_max_target;
241 
242 	device_t		xbdev;
243 	uint8_t			mac[ETHER_ADDR_LEN];
244 
245 	int			xn_if_flags;
246 
247 	struct ifmedia		sc_media;
248 
249 	bool			xn_resume;
250 };
251 
252 struct netfront_rx_info {
253 	struct netif_rx_response rx;
254 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
255 };
256 
257 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
258 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
259 
260 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
261 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
262 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
263 
264 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
265 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
266 
267 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
268 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
269 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
270 
271 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
272 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
273 #define netfront_carrier_ok(netif)	((netif)->carrier)
274 
275 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
276 
277 static inline void
278 add_id_to_freelist(struct mbuf **list, uintptr_t id)
279 {
280 
281 	KASSERT(id != 0,
282 		("%s: the head item (0) must always be free.", __func__));
283 	list[id] = list[0];
284 	list[0]  = (struct mbuf *)id;
285 }
286 
287 static inline unsigned short
288 get_id_from_freelist(struct mbuf **list)
289 {
290 	uintptr_t id;
291 
292 	id = (uintptr_t)list[0];
293 	KASSERT(id != 0,
294 		("%s: the head item (0) must always remain free.", __func__));
295 	list[0] = list[id];
296 	return (id);
297 }
298 
299 static inline int
300 xn_rxidx(RING_IDX idx)
301 {
302 
303 	return idx & (NET_RX_RING_SIZE - 1);
304 }
305 
306 static inline struct mbuf *
307 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
308 {
309 	int i;
310 	struct mbuf *m;
311 
312 	i = xn_rxidx(ri);
313 	m = rxq->mbufs[i];
314 	rxq->mbufs[i] = NULL;
315 	return (m);
316 }
317 
318 static inline grant_ref_t
319 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
320 {
321 	int i = xn_rxidx(ri);
322 	grant_ref_t ref = rxq->grant_ref[i];
323 
324 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
325 	rxq->grant_ref[i] = GRANT_REF_INVALID;
326 	return (ref);
327 }
328 
329 #define IPRINTK(fmt, args...) \
330     printf("[XEN] " fmt, ##args)
331 #ifdef INVARIANTS
332 #define WPRINTK(fmt, args...) \
333     printf("[XEN] " fmt, ##args)
334 #else
335 #define WPRINTK(fmt, args...)
336 #endif
337 #ifdef DEBUG
338 #define DPRINTK(fmt, args...) \
339     printf("[XEN] %s: " fmt, __func__, ##args)
340 #else
341 #define DPRINTK(fmt, args...)
342 #endif
343 
344 /**
345  * Read the 'mac' node at the given device's node in the store, and parse that
346  * as colon-separated octets, placing result the given mac array.  mac must be
347  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
348  * Return 0 on success, or errno on error.
349  */
350 static int
351 xen_net_read_mac(device_t dev, uint8_t mac[])
352 {
353 	int error, i;
354 	char *s, *e, *macstr;
355 	const char *path;
356 
357 	path = xenbus_get_node(dev);
358 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
359 	if (error == ENOENT) {
360 		/*
361 		 * Deal with missing mac XenStore nodes on devices with
362 		 * HVM emulation (the 'ioemu' configuration attribute)
363 		 * enabled.
364 		 *
365 		 * The HVM emulator may execute in a stub device model
366 		 * domain which lacks the permission, only given to Dom0,
367 		 * to update the guest's XenStore tree.  For this reason,
368 		 * the HVM emulator doesn't even attempt to write the
369 		 * front-side mac node, even when operating in Dom0.
370 		 * However, there should always be a mac listed in the
371 		 * backend tree.  Fallback to this version if our query
372 		 * of the front side XenStore location doesn't find
373 		 * anything.
374 		 */
375 		path = xenbus_get_otherend_path(dev);
376 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
377 	}
378 	if (error != 0) {
379 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
380 		return (error);
381 	}
382 
383 	s = macstr;
384 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
385 		mac[i] = strtoul(s, &e, 16);
386 		if (s == e || (e[0] != ':' && e[0] != 0)) {
387 			free(macstr, M_XENBUS);
388 			return (ENOENT);
389 		}
390 		s = &e[1];
391 	}
392 	free(macstr, M_XENBUS);
393 	return (0);
394 }
395 
396 /**
397  * Entry point to this code when a new device is created.  Allocate the basic
398  * structures and the ring buffers for communication with the backend, and
399  * inform the backend of the appropriate details for those.  Switch to
400  * Connected state.
401  */
402 static int
403 netfront_probe(device_t dev)
404 {
405 
406 	if (xen_hvm_domain() && xen_disable_pv_nics != 0)
407 		return (ENXIO);
408 
409 	if (!strcmp(xenbus_get_type(dev), "vif")) {
410 		device_set_desc(dev, "Virtual Network Interface");
411 		return (0);
412 	}
413 
414 	return (ENXIO);
415 }
416 
417 static int
418 netfront_attach(device_t dev)
419 {
420 	int err;
421 
422 	err = create_netdev(dev);
423 	if (err != 0) {
424 		xenbus_dev_fatal(dev, err, "creating netdev");
425 		return (err);
426 	}
427 
428 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
429 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
430 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
431 	    &xn_enable_lro, 0, "Large Receive Offload");
432 
433 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
434 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
435 	    OID_AUTO, "num_queues", CTLFLAG_RD,
436 	    &xn_num_queues, "Number of pairs of queues");
437 
438 	return (0);
439 }
440 
441 static int
442 netfront_suspend(device_t dev)
443 {
444 	struct netfront_info *np = device_get_softc(dev);
445 	u_int i;
446 
447 	for (i = 0; i < np->num_queues; i++) {
448 		XN_RX_LOCK(&np->rxq[i]);
449 		XN_TX_LOCK(&np->txq[i]);
450 	}
451 	netfront_carrier_off(np);
452 	for (i = 0; i < np->num_queues; i++) {
453 		XN_RX_UNLOCK(&np->rxq[i]);
454 		XN_TX_UNLOCK(&np->txq[i]);
455 	}
456 	return (0);
457 }
458 
459 /**
460  * We are reconnecting to the backend, due to a suspend/resume, or a backend
461  * driver restart.  We tear down our netif structure and recreate it, but
462  * leave the device-layer structures intact so that this is transparent to the
463  * rest of the kernel.
464  */
465 static int
466 netfront_resume(device_t dev)
467 {
468 	struct netfront_info *info = device_get_softc(dev);
469 
470 	info->xn_resume = true;
471 	netif_disconnect_backend(info);
472 	return (0);
473 }
474 
475 static int
476 write_queue_xenstore_keys(device_t dev,
477     struct netfront_rxq *rxq,
478     struct netfront_txq *txq,
479     struct xs_transaction *xst, bool hierarchy)
480 {
481 	int err;
482 	const char *message;
483 	const char *node = xenbus_get_node(dev);
484 	char *path;
485 	size_t path_size;
486 
487 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
488 	/* Split event channel support is not yet there. */
489 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
490 	    ("Split event channels are not supported"));
491 
492 	if (hierarchy) {
493 		path_size = strlen(node) + 10;
494 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
495 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
496 	} else {
497 		path_size = strlen(node) + 1;
498 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
499 		snprintf(path, path_size, "%s", node);
500 	}
501 
502 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
503 	if (err != 0) {
504 		message = "writing tx ring-ref";
505 		goto error;
506 	}
507 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
508 	if (err != 0) {
509 		message = "writing rx ring-ref";
510 		goto error;
511 	}
512 	err = xs_printf(*xst, path, "event-channel", "%u",
513 	    xen_intr_port(rxq->xen_intr_handle));
514 	if (err != 0) {
515 		message = "writing event-channel";
516 		goto error;
517 	}
518 
519 	free(path, M_DEVBUF);
520 
521 	return (0);
522 
523 error:
524 	free(path, M_DEVBUF);
525 	xenbus_dev_fatal(dev, err, "%s", message);
526 
527 	return (err);
528 }
529 
530 /* Common code used when first setting up, and when resuming. */
531 static int
532 talk_to_backend(device_t dev, struct netfront_info *info)
533 {
534 	const char *message;
535 	struct xs_transaction xst;
536 	const char *node = xenbus_get_node(dev);
537 	int err;
538 	unsigned long num_queues, max_queues = 0;
539 	unsigned int i;
540 
541 	err = xen_net_read_mac(dev, info->mac);
542 	if (err != 0) {
543 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
544 		goto out;
545 	}
546 
547 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
548 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
549 	if (err != 0)
550 		max_queues = 1;
551 	num_queues = xn_num_queues;
552 	if (num_queues > max_queues)
553 		num_queues = max_queues;
554 
555 	err = setup_device(dev, info, num_queues);
556 	if (err != 0)
557 		goto out;
558 
559  again:
560 	err = xs_transaction_start(&xst);
561 	if (err != 0) {
562 		xenbus_dev_fatal(dev, err, "starting transaction");
563 		goto free;
564 	}
565 
566 	if (info->num_queues == 1) {
567 		err = write_queue_xenstore_keys(dev, &info->rxq[0],
568 		    &info->txq[0], &xst, false);
569 		if (err != 0)
570 			goto abort_transaction_no_def_error;
571 	} else {
572 		err = xs_printf(xst, node, "multi-queue-num-queues",
573 		    "%u", info->num_queues);
574 		if (err != 0) {
575 			message = "writing multi-queue-num-queues";
576 			goto abort_transaction;
577 		}
578 
579 		for (i = 0; i < info->num_queues; i++) {
580 			err = write_queue_xenstore_keys(dev, &info->rxq[i],
581 			    &info->txq[i], &xst, true);
582 			if (err != 0)
583 				goto abort_transaction_no_def_error;
584 		}
585 	}
586 
587 	err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
588 	if (err != 0) {
589 		message = "writing request-rx-copy";
590 		goto abort_transaction;
591 	}
592 	err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
593 	if (err != 0) {
594 		message = "writing feature-rx-notify";
595 		goto abort_transaction;
596 	}
597 	err = xs_printf(xst, node, "feature-sg", "%d", 1);
598 	if (err != 0) {
599 		message = "writing feature-sg";
600 		goto abort_transaction;
601 	}
602 	err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
603 	if (err != 0) {
604 		message = "writing feature-gso-tcpv4";
605 		goto abort_transaction;
606 	}
607 
608 	err = xs_transaction_end(xst, 0);
609 	if (err != 0) {
610 		if (err == EAGAIN)
611 			goto again;
612 		xenbus_dev_fatal(dev, err, "completing transaction");
613 		goto free;
614 	}
615 
616 	return 0;
617 
618  abort_transaction:
619 	xenbus_dev_fatal(dev, err, "%s", message);
620  abort_transaction_no_def_error:
621 	xs_transaction_end(xst, 1);
622  free:
623 	netif_free(info);
624  out:
625 	return (err);
626 }
627 
628 static void
629 xn_rxq_tq_intr(void *xrxq, int pending)
630 {
631 	struct netfront_rxq *rxq = xrxq;
632 
633 	XN_RX_LOCK(rxq);
634 	xn_rxeof(rxq);
635 	XN_RX_UNLOCK(rxq);
636 }
637 
638 static void
639 xn_txq_start(struct netfront_txq *txq)
640 {
641 	struct netfront_info *np = txq->info;
642 	struct ifnet *ifp = np->xn_ifp;
643 
644 	XN_TX_LOCK_ASSERT(txq);
645 	if (!drbr_empty(ifp, txq->br))
646 		xn_txq_mq_start_locked(txq, NULL);
647 }
648 
649 static void
650 xn_txq_tq_intr(void *xtxq, int pending)
651 {
652 	struct netfront_txq *txq = xtxq;
653 
654 	XN_TX_LOCK(txq);
655 	if (RING_HAS_UNCONSUMED_RESPONSES(&txq->ring))
656 		xn_txeof(txq);
657 	xn_txq_start(txq);
658 	XN_TX_UNLOCK(txq);
659 }
660 
661 static void
662 xn_txq_tq_deferred(void *xtxq, int pending)
663 {
664 	struct netfront_txq *txq = xtxq;
665 
666 	XN_TX_LOCK(txq);
667 	xn_txq_start(txq);
668 	XN_TX_UNLOCK(txq);
669 }
670 
671 static void
672 disconnect_rxq(struct netfront_rxq *rxq)
673 {
674 
675 	xn_release_rx_bufs(rxq);
676 	gnttab_free_grant_references(rxq->gref_head);
677 	gnttab_end_foreign_access_ref(rxq->ring_ref);
678 	/*
679 	 * No split event channel support at the moment, handle will
680 	 * be unbound in tx. So no need to call xen_intr_unbind here,
681 	 * but we do want to reset the handler to 0.
682 	 */
683 	rxq->xen_intr_handle = 0;
684 }
685 
686 static void
687 destroy_rxq(struct netfront_rxq *rxq)
688 {
689 
690 	free(rxq->ring.sring, M_DEVBUF);
691 	taskqueue_drain_all(rxq->tq);
692 	taskqueue_free(rxq->tq);
693 }
694 
695 static void
696 destroy_rxqs(struct netfront_info *np)
697 {
698 	int i;
699 
700 	for (i = 0; i < np->num_queues; i++)
701 		destroy_rxq(&np->rxq[i]);
702 
703 	free(np->rxq, M_DEVBUF);
704 	np->rxq = NULL;
705 }
706 
707 static int
708 setup_rxqs(device_t dev, struct netfront_info *info,
709 	   unsigned long num_queues)
710 {
711 	int q, i;
712 	int error;
713 	netif_rx_sring_t *rxs;
714 	struct netfront_rxq *rxq;
715 
716 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
717 	    M_DEVBUF, M_WAITOK|M_ZERO);
718 
719 	for (q = 0; q < num_queues; q++) {
720 		rxq = &info->rxq[q];
721 
722 		rxq->id = q;
723 		rxq->info = info;
724 		rxq->target = RX_MIN_TARGET;
725 		rxq->ring_ref = GRANT_REF_INVALID;
726 		rxq->ring.sring = NULL;
727 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
728 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
729 		    MTX_DEF);
730 
731 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
732 			rxq->mbufs[i] = NULL;
733 			rxq->grant_ref[i] = GRANT_REF_INVALID;
734 		}
735 
736 		mbufq_init(&rxq->batch, INT_MAX);
737 
738 		/* Start resources allocation */
739 
740 		if (gnttab_alloc_grant_references(RX_MAX_TARGET,
741 		    &rxq->gref_head) != 0) {
742 			device_printf(dev, "allocating rx gref");
743 			error = ENOMEM;
744 			goto fail;
745 		}
746 
747 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
748 		    M_WAITOK|M_ZERO);
749 		SHARED_RING_INIT(rxs);
750 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
751 
752 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
753 		    &rxq->ring_ref);
754 		if (error != 0) {
755 			device_printf(dev, "granting rx ring page");
756 			goto fail_grant_ring;
757 		}
758 
759 		TASK_INIT(&rxq->intrtask, 0, xn_rxq_tq_intr, rxq);
760 		rxq->tq = taskqueue_create_fast(rxq->name, M_WAITOK,
761 		    taskqueue_thread_enqueue, &rxq->tq);
762 
763 		error = taskqueue_start_threads(&rxq->tq, 1, PI_NET,
764 		    "%s rxq %d", device_get_nameunit(dev), rxq->id);
765 		if (error != 0) {
766 			device_printf(dev, "failed to start rx taskq %d\n",
767 			    rxq->id);
768 			goto fail_start_thread;
769 		}
770 	}
771 
772 	return (0);
773 
774 fail_start_thread:
775 	gnttab_end_foreign_access_ref(rxq->ring_ref);
776 	taskqueue_drain_all(rxq->tq);
777 	taskqueue_free(rxq->tq);
778 fail_grant_ring:
779 	gnttab_free_grant_references(rxq->gref_head);
780 	free(rxq->ring.sring, M_DEVBUF);
781 fail:
782 	for (; q >= 0; q--) {
783 		disconnect_rxq(&info->rxq[q]);
784 		destroy_rxq(&info->rxq[q]);
785 	}
786 
787 	free(info->rxq, M_DEVBUF);
788 	return (error);
789 }
790 
791 static void
792 disconnect_txq(struct netfront_txq *txq)
793 {
794 
795 	xn_release_tx_bufs(txq);
796 	gnttab_free_grant_references(txq->gref_head);
797 	gnttab_end_foreign_access_ref(txq->ring_ref);
798 	xen_intr_unbind(&txq->xen_intr_handle);
799 }
800 
801 static void
802 destroy_txq(struct netfront_txq *txq)
803 {
804 
805 	free(txq->ring.sring, M_DEVBUF);
806 	buf_ring_free(txq->br, M_DEVBUF);
807 	taskqueue_drain_all(txq->tq);
808 	taskqueue_free(txq->tq);
809 }
810 
811 static void
812 destroy_txqs(struct netfront_info *np)
813 {
814 	int i;
815 
816 	for (i = 0; i < np->num_queues; i++)
817 		destroy_txq(&np->txq[i]);
818 
819 	free(np->txq, M_DEVBUF);
820 	np->txq = NULL;
821 }
822 
823 static int
824 setup_txqs(device_t dev, struct netfront_info *info,
825 	   unsigned long num_queues)
826 {
827 	int q, i;
828 	int error;
829 	netif_tx_sring_t *txs;
830 	struct netfront_txq *txq;
831 
832 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
833 	    M_DEVBUF, M_WAITOK|M_ZERO);
834 
835 	for (q = 0; q < num_queues; q++) {
836 		txq = &info->txq[q];
837 
838 		txq->id = q;
839 		txq->info = info;
840 
841 		txq->ring_ref = GRANT_REF_INVALID;
842 		txq->ring.sring = NULL;
843 
844 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
845 
846 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
847 		    MTX_DEF);
848 
849 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
850 			txq->mbufs[i] = (void *) ((u_long) i+1);
851 			txq->grant_ref[i] = GRANT_REF_INVALID;
852 		}
853 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
854 
855 		/* Start resources allocation. */
856 
857 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
858 		    &txq->gref_head) != 0) {
859 			device_printf(dev, "failed to allocate tx grant refs\n");
860 			error = ENOMEM;
861 			goto fail;
862 		}
863 
864 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
865 		    M_WAITOK|M_ZERO);
866 		SHARED_RING_INIT(txs);
867 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
868 
869 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
870 		    &txq->ring_ref);
871 		if (error != 0) {
872 			device_printf(dev, "failed to grant tx ring\n");
873 			goto fail_grant_ring;
874 		}
875 
876 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
877 		    M_WAITOK, &txq->lock);
878 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
879 		TASK_INIT(&txq->intrtask, 0, xn_txq_tq_intr, txq);
880 
881 		txq->tq = taskqueue_create_fast(txq->name, M_WAITOK,
882 		    taskqueue_thread_enqueue, &txq->tq);
883 
884 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
885 		    "%s txq %d", device_get_nameunit(dev), txq->id);
886 		if (error != 0) {
887 			device_printf(dev, "failed to start tx taskq %d\n",
888 			    txq->id);
889 			goto fail_start_thread;
890 		}
891 
892 		error = xen_intr_alloc_and_bind_local_port(dev,
893 			    xenbus_get_otherend_id(dev), xn_intr, /* handler */ NULL,
894 			    &info->txq[q],
895 			    INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
896 			    &txq->xen_intr_handle);
897 
898 		if (error != 0) {
899 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
900 			goto fail_bind_port;
901 		}
902 	}
903 
904 	return (0);
905 
906 fail_bind_port:
907 	taskqueue_drain_all(txq->tq);
908 fail_start_thread:
909 	gnttab_free_grant_references(txq->gref_head);
910 	free(txq->ring.sring, M_DEVBUF);
911 	gnttab_end_foreign_access_ref(txq->ring_ref);
912 	buf_ring_free(txq->br, M_DEVBUF);
913 	taskqueue_free(txq->tq);
914 fail_grant_ring:
915 	gnttab_free_grant_references(txq->gref_head);
916 	free(txq->ring.sring, M_DEVBUF);
917 fail:
918 	for (; q >= 0; q--) {
919 		disconnect_txq(&info->txq[q]);
920 		destroy_txq(&info->txq[q]);
921 	}
922 
923 	free(info->txq, M_DEVBUF);
924 	return (error);
925 }
926 
927 static int
928 setup_device(device_t dev, struct netfront_info *info,
929     unsigned long num_queues)
930 {
931 	int error;
932 	int q;
933 
934 	if (info->txq)
935 		destroy_txqs(info);
936 
937 	if (info->rxq)
938 		destroy_rxqs(info);
939 
940 	info->num_queues = 0;
941 
942 	error = setup_rxqs(dev, info, num_queues);
943 	if (error != 0)
944 		goto out;
945 	error = setup_txqs(dev, info, num_queues);
946 	if (error != 0)
947 		goto out;
948 
949 	info->num_queues = num_queues;
950 
951 	/* No split event channel at the moment. */
952 	for (q = 0; q < num_queues; q++)
953 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
954 
955 	return (0);
956 
957 out:
958 	KASSERT(error != 0, ("Error path taken without providing an error code"));
959 	return (error);
960 }
961 
962 #ifdef INET
963 /**
964  * If this interface has an ipv4 address, send an arp for it. This
965  * helps to get the network going again after migrating hosts.
966  */
967 static void
968 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
969 {
970 	struct ifnet *ifp;
971 	struct ifaddr *ifa;
972 
973 	ifp = info->xn_ifp;
974 	TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
975 		if (ifa->ifa_addr->sa_family == AF_INET) {
976 			arp_ifinit(ifp, ifa);
977 		}
978 	}
979 }
980 #endif
981 
982 /**
983  * Callback received when the backend's state changes.
984  */
985 static void
986 netfront_backend_changed(device_t dev, XenbusState newstate)
987 {
988 	struct netfront_info *sc = device_get_softc(dev);
989 
990 	DPRINTK("newstate=%d\n", newstate);
991 
992 	switch (newstate) {
993 	case XenbusStateInitialising:
994 	case XenbusStateInitialised:
995 	case XenbusStateUnknown:
996 	case XenbusStateClosed:
997 	case XenbusStateReconfigured:
998 	case XenbusStateReconfiguring:
999 		break;
1000 	case XenbusStateInitWait:
1001 		if (xenbus_get_state(dev) != XenbusStateInitialising)
1002 			break;
1003 		if (xn_connect(sc) != 0)
1004 			break;
1005 		xenbus_set_state(dev, XenbusStateConnected);
1006 		break;
1007 	case XenbusStateClosing:
1008 		xenbus_set_state(dev, XenbusStateClosed);
1009 		break;
1010 	case XenbusStateConnected:
1011 #ifdef INET
1012 		netfront_send_fake_arp(dev, sc);
1013 #endif
1014 		break;
1015 	}
1016 }
1017 
1018 /**
1019  * \brief Verify that there is sufficient space in the Tx ring
1020  *        buffer for a maximally sized request to be enqueued.
1021  *
1022  * A transmit request requires a transmit descriptor for each packet
1023  * fragment, plus up to 2 entries for "options" (e.g. TSO).
1024  */
1025 static inline int
1026 xn_tx_slot_available(struct netfront_txq *txq)
1027 {
1028 
1029 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1030 }
1031 
1032 static void
1033 xn_release_tx_bufs(struct netfront_txq *txq)
1034 {
1035 	int i;
1036 
1037 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1038 		struct mbuf *m;
1039 
1040 		m = txq->mbufs[i];
1041 
1042 		/*
1043 		 * We assume that no kernel addresses are
1044 		 * less than NET_TX_RING_SIZE.  Any entry
1045 		 * in the table that is below this number
1046 		 * must be an index from free-list tracking.
1047 		 */
1048 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1049 			continue;
1050 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1051 		gnttab_release_grant_reference(&txq->gref_head,
1052 		    txq->grant_ref[i]);
1053 		txq->grant_ref[i] = GRANT_REF_INVALID;
1054 		add_id_to_freelist(txq->mbufs, i);
1055 		txq->mbufs_cnt--;
1056 		if (txq->mbufs_cnt < 0) {
1057 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1058 		}
1059 		m_free(m);
1060 	}
1061 }
1062 
1063 static void
1064 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1065 {
1066 	struct netfront_info *np = rxq->info;
1067 	int otherend_id = xenbus_get_otherend_id(np->xbdev);
1068 	unsigned short id;
1069 	struct mbuf *m_new;
1070 	int i, batch_target, notify;
1071 	RING_IDX req_prod;
1072 	grant_ref_t ref;
1073 	netif_rx_request_t *req;
1074 	vm_offset_t vaddr;
1075 	u_long pfn;
1076 
1077 	req_prod = rxq->ring.req_prod_pvt;
1078 
1079 	if (__predict_false(np->carrier == 0))
1080 		return;
1081 
1082 	/*
1083 	 * Allocate mbufs greedily, even though we batch updates to the
1084 	 * receive ring. This creates a less bursty demand on the memory
1085 	 * allocator, and so should reduce the chance of failed allocation
1086 	 * requests both for ourself and for other kernel subsystems.
1087 	 *
1088 	 * Here we attempt to maintain rx_target buffers in flight, counting
1089 	 * buffers that we have yet to process in the receive ring.
1090 	 */
1091 	batch_target = rxq->target - (req_prod - rxq->ring.rsp_cons);
1092 	for (i = mbufq_len(&rxq->batch); i < batch_target; i++) {
1093 		m_new = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1094 		if (m_new == NULL) {
1095 			if (i != 0)
1096 				goto refill;
1097 			/* XXX set timer */
1098 			break;
1099 		}
1100 		m_new->m_len = m_new->m_pkthdr.len = MJUMPAGESIZE;
1101 
1102 		/* queue the mbufs allocated */
1103 		mbufq_enqueue(&rxq->batch, m_new);
1104 	}
1105 
1106 	/*
1107 	 * If we've allocated at least half of our target number of entries,
1108 	 * submit them to the backend - we have enough to make the overhead
1109 	 * of submission worthwhile.  Otherwise wait for more mbufs and
1110 	 * request entries to become available.
1111 	 */
1112 	if (i < (rxq->target/2)) {
1113 		if (req_prod > rxq->ring.sring->req_prod)
1114 			goto push;
1115 		return;
1116 	}
1117 
1118 	/*
1119 	 * Double floating fill target if we risked having the backend
1120 	 * run out of empty buffers for receive traffic.  We define "running
1121 	 * low" as having less than a fourth of our target buffers free
1122 	 * at the time we refilled the queue.
1123 	 */
1124 	if ((req_prod - rxq->ring.sring->rsp_prod) < (rxq->target / 4)) {
1125 		rxq->target *= 2;
1126 		if (rxq->target > np->rx_max_target)
1127 			rxq->target = np->rx_max_target;
1128 	}
1129 
1130 refill:
1131 	for (i = 0; ; i++) {
1132 		if ((m_new = mbufq_dequeue(&rxq->batch)) == NULL)
1133 			break;
1134 
1135 		m_new->m_ext.ext_arg1 = (vm_paddr_t *)(uintptr_t)(
1136 				vtophys(m_new->m_ext.ext_buf) >> PAGE_SHIFT);
1137 
1138 		id = xn_rxidx(req_prod + i);
1139 
1140 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1141 		rxq->mbufs[id] = m_new;
1142 
1143 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1144 		KASSERT(ref != GNTTAB_LIST_END,
1145 			("reserved grant references exhuasted"));
1146 		rxq->grant_ref[id] = ref;
1147 
1148 		vaddr = mtod(m_new, vm_offset_t);
1149 		pfn = vtophys(vaddr) >> PAGE_SHIFT;
1150 		req = RING_GET_REQUEST(&rxq->ring, req_prod + i);
1151 
1152 		gnttab_grant_foreign_access_ref(ref, otherend_id, pfn, 0);
1153 		req->id = id;
1154 		req->gref = ref;
1155 
1156 		rxq->pfn_array[i] =
1157 		    vtophys(mtod(m_new,vm_offset_t)) >> PAGE_SHIFT;
1158 	}
1159 
1160 	KASSERT(i, ("no mbufs processed")); /* should have returned earlier */
1161 	KASSERT(mbufq_len(&rxq->batch) == 0, ("not all mbufs processed"));
1162 	/*
1163 	 * We may have allocated buffers which have entries outstanding
1164 	 * in the page * update queue -- make sure we flush those first!
1165 	 */
1166 	wmb();
1167 
1168 	/* Above is a suitable barrier to ensure backend will see requests. */
1169 	rxq->ring.req_prod_pvt = req_prod + i;
1170 push:
1171 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1172 	if (notify)
1173 		xen_intr_signal(rxq->xen_intr_handle);
1174 }
1175 
1176 static void
1177 xn_release_rx_bufs(struct netfront_rxq *rxq)
1178 {
1179 	int i,  ref;
1180 	struct mbuf *m;
1181 
1182 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1183 		m = rxq->mbufs[i];
1184 
1185 		if (m == NULL)
1186 			continue;
1187 
1188 		ref = rxq->grant_ref[i];
1189 		if (ref == GRANT_REF_INVALID)
1190 			continue;
1191 
1192 		gnttab_end_foreign_access_ref(ref);
1193 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1194 		rxq->mbufs[i] = NULL;
1195 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1196 		m_freem(m);
1197 	}
1198 }
1199 
1200 static void
1201 xn_rxeof(struct netfront_rxq *rxq)
1202 {
1203 	struct ifnet *ifp;
1204 	struct netfront_info *np = rxq->info;
1205 #if (defined(INET) || defined(INET6))
1206 	struct lro_ctrl *lro = &rxq->lro;
1207 	struct lro_entry *queued;
1208 #endif
1209 	struct netfront_rx_info rinfo;
1210 	struct netif_rx_response *rx = &rinfo.rx;
1211 	struct netif_extra_info *extras = rinfo.extras;
1212 	RING_IDX i, rp;
1213 	struct mbuf *m;
1214 	struct mbufq mbufq_rxq, mbufq_errq;
1215 	int err, work_to_do;
1216 
1217 	do {
1218 		XN_RX_LOCK_ASSERT(rxq);
1219 		if (!netfront_carrier_ok(np))
1220 			return;
1221 
1222 		/* XXX: there should be some sane limit. */
1223 		mbufq_init(&mbufq_errq, INT_MAX);
1224 		mbufq_init(&mbufq_rxq, INT_MAX);
1225 
1226 		ifp = np->xn_ifp;
1227 
1228 		rp = rxq->ring.sring->rsp_prod;
1229 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1230 
1231 		i = rxq->ring.rsp_cons;
1232 		while ((i != rp)) {
1233 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1234 			memset(extras, 0, sizeof(rinfo.extras));
1235 
1236 			m = NULL;
1237 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1238 
1239 			if (__predict_false(err)) {
1240 				if (m)
1241 					(void )mbufq_enqueue(&mbufq_errq, m);
1242 				rxq->stats.rx_errors++;
1243 				continue;
1244 			}
1245 
1246 			m->m_pkthdr.rcvif = ifp;
1247 			if ( rx->flags & NETRXF_data_validated ) {
1248 				/* Tell the stack the checksums are okay */
1249 				/*
1250 				 * XXX this isn't necessarily the case - need to add
1251 				 * check
1252 				 */
1253 
1254 				m->m_pkthdr.csum_flags |=
1255 					(CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID
1256 					    | CSUM_PSEUDO_HDR);
1257 				m->m_pkthdr.csum_data = 0xffff;
1258 			}
1259 
1260 			rxq->stats.rx_packets++;
1261 			rxq->stats.rx_bytes += m->m_pkthdr.len;
1262 
1263 			(void )mbufq_enqueue(&mbufq_rxq, m);
1264 			rxq->ring.rsp_cons = i;
1265 		}
1266 
1267 		mbufq_drain(&mbufq_errq);
1268 
1269 		/*
1270 		 * Process all the mbufs after the remapping is complete.
1271 		 * Break the mbuf chain first though.
1272 		 */
1273 		while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1274 			if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1275 
1276 			/* XXX: Do we really need to drop the rx lock? */
1277 			XN_RX_UNLOCK(rxq);
1278 #if (defined(INET) || defined(INET6))
1279 			/* Use LRO if possible */
1280 			if ((ifp->if_capenable & IFCAP_LRO) == 0 ||
1281 			    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1282 				/*
1283 				 * If LRO fails, pass up to the stack
1284 				 * directly.
1285 				 */
1286 				(*ifp->if_input)(ifp, m);
1287 			}
1288 #else
1289 			(*ifp->if_input)(ifp, m);
1290 #endif
1291 
1292 			XN_RX_LOCK(rxq);
1293 		}
1294 
1295 		rxq->ring.rsp_cons = i;
1296 
1297 #if (defined(INET) || defined(INET6))
1298 		/*
1299 		 * Flush any outstanding LRO work
1300 		 */
1301 		while (!SLIST_EMPTY(&lro->lro_active)) {
1302 			queued = SLIST_FIRST(&lro->lro_active);
1303 			SLIST_REMOVE_HEAD(&lro->lro_active, next);
1304 			tcp_lro_flush(lro, queued);
1305 		}
1306 #endif
1307 
1308 		xn_alloc_rx_buffers(rxq);
1309 
1310 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1311 	} while (work_to_do);
1312 }
1313 
1314 static void
1315 xn_txeof(struct netfront_txq *txq)
1316 {
1317 	RING_IDX i, prod;
1318 	unsigned short id;
1319 	struct ifnet *ifp;
1320 	netif_tx_response_t *txr;
1321 	struct mbuf *m;
1322 	struct netfront_info *np = txq->info;
1323 
1324 	XN_TX_LOCK_ASSERT(txq);
1325 
1326 	if (!netfront_carrier_ok(np))
1327 		return;
1328 
1329 	ifp = np->xn_ifp;
1330 
1331 	do {
1332 		prod = txq->ring.sring->rsp_prod;
1333 		rmb(); /* Ensure we see responses up to 'rp'. */
1334 
1335 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1336 			txr = RING_GET_RESPONSE(&txq->ring, i);
1337 			if (txr->status == NETIF_RSP_NULL)
1338 				continue;
1339 
1340 			if (txr->status != NETIF_RSP_OKAY) {
1341 				printf("%s: WARNING: response is %d!\n",
1342 				       __func__, txr->status);
1343 			}
1344 			id = txr->id;
1345 			m = txq->mbufs[id];
1346 			KASSERT(m != NULL, ("mbuf not found in chain"));
1347 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1348 				("mbuf already on the free list, but we're "
1349 				"trying to free it again!"));
1350 			M_ASSERTVALID(m);
1351 
1352 			/*
1353 			 * Increment packet count if this is the last
1354 			 * mbuf of the chain.
1355 			 */
1356 			if (!m->m_next)
1357 				if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1358 			if (__predict_false(gnttab_query_foreign_access(
1359 			    txq->grant_ref[id]) != 0)) {
1360 				panic("%s: grant id %u still in use by the "
1361 				    "backend", __func__, id);
1362 			}
1363 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1364 			gnttab_release_grant_reference(
1365 				&txq->gref_head, txq->grant_ref[id]);
1366 			txq->grant_ref[id] = GRANT_REF_INVALID;
1367 
1368 			txq->mbufs[id] = NULL;
1369 			add_id_to_freelist(txq->mbufs, id);
1370 			txq->mbufs_cnt--;
1371 			m_free(m);
1372 			/* Only mark the txq active if we've freed up at least one slot to try */
1373 			ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1374 		}
1375 		txq->ring.rsp_cons = prod;
1376 
1377 		/*
1378 		 * Set a new event, then check for race with update of
1379 		 * tx_cons. Note that it is essential to schedule a
1380 		 * callback, no matter how few buffers are pending. Even if
1381 		 * there is space in the transmit ring, higher layers may
1382 		 * be blocked because too much data is outstanding: in such
1383 		 * cases notification from Xen is likely to be the only kick
1384 		 * that we'll get.
1385 		 */
1386 		txq->ring.sring->rsp_event =
1387 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1388 
1389 		mb();
1390 	} while (prod != txq->ring.sring->rsp_prod);
1391 
1392 	if (txq->full &&
1393 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1394 		txq->full = false;
1395 		taskqueue_enqueue(txq->tq, &txq->intrtask);
1396 	}
1397 }
1398 
1399 
1400 static void
1401 xn_rxq_intr(void *xrxq)
1402 {
1403 	struct netfront_rxq *rxq = xrxq;
1404 
1405 	taskqueue_enqueue_fast(rxq->tq, &rxq->intrtask);
1406 }
1407 
1408 static void
1409 xn_txq_intr(void *xtxq)
1410 {
1411 	struct netfront_txq *txq = xtxq;
1412 
1413 	taskqueue_enqueue_fast(txq->tq, &txq->intrtask);
1414 }
1415 
1416 static int
1417 xn_intr(void *xsc)
1418 {
1419 	struct netfront_txq *txq = xsc;
1420 	struct netfront_info *np = txq->info;
1421 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1422 
1423 	/* kick both tx and rx */
1424 	xn_rxq_intr(rxq);
1425 	xn_txq_intr(txq);
1426 
1427 	return (FILTER_HANDLED);
1428 }
1429 
1430 static void
1431 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1432     grant_ref_t ref)
1433 {
1434 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1435 
1436 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1437 	rxq->mbufs[new] = m;
1438 	rxq->grant_ref[new] = ref;
1439 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1440 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1441 	rxq->ring.req_prod_pvt++;
1442 }
1443 
1444 static int
1445 xn_get_extras(struct netfront_rxq *rxq,
1446     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1447 {
1448 	struct netif_extra_info *extra;
1449 
1450 	int err = 0;
1451 
1452 	do {
1453 		struct mbuf *m;
1454 		grant_ref_t ref;
1455 
1456 		if (__predict_false(*cons + 1 == rp)) {
1457 			err = EINVAL;
1458 			break;
1459 		}
1460 
1461 		extra = (struct netif_extra_info *)
1462 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1463 
1464 		if (__predict_false(!extra->type ||
1465 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1466 			err = EINVAL;
1467 		} else {
1468 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1469 		}
1470 
1471 		m = xn_get_rx_mbuf(rxq, *cons);
1472 		ref = xn_get_rx_ref(rxq,  *cons);
1473 		xn_move_rx_slot(rxq, m, ref);
1474 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1475 
1476 	return err;
1477 }
1478 
1479 static int
1480 xn_get_responses(struct netfront_rxq *rxq,
1481     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1482     struct mbuf  **list)
1483 {
1484 	struct netif_rx_response *rx = &rinfo->rx;
1485 	struct netif_extra_info *extras = rinfo->extras;
1486 	struct mbuf *m, *m0, *m_prev;
1487 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1488 	RING_IDX ref_cons = *cons;
1489 	int frags = 1;
1490 	int err = 0;
1491 	u_long ret;
1492 
1493 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1494 
1495 	if (rx->flags & NETRXF_extra_info) {
1496 		err = xn_get_extras(rxq, extras, rp, cons);
1497 	}
1498 
1499 	if (m0 != NULL) {
1500 		m0->m_pkthdr.len = 0;
1501 		m0->m_next = NULL;
1502 	}
1503 
1504 	for (;;) {
1505 #if 0
1506 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1507 			rx->status, rx->offset, frags);
1508 #endif
1509 		if (__predict_false(rx->status < 0 ||
1510 			rx->offset + rx->status > PAGE_SIZE)) {
1511 
1512 			xn_move_rx_slot(rxq, m, ref);
1513 			if (m0 == m)
1514 				m0 = NULL;
1515 			m = NULL;
1516 			err = EINVAL;
1517 			goto next_skip_queue;
1518 		}
1519 
1520 		/*
1521 		 * This definitely indicates a bug, either in this driver or in
1522 		 * the backend driver. In future this should flag the bad
1523 		 * situation to the system controller to reboot the backed.
1524 		 */
1525 		if (ref == GRANT_REF_INVALID) {
1526 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1527 			err = EINVAL;
1528 			goto next;
1529 		}
1530 
1531 		ret = gnttab_end_foreign_access_ref(ref);
1532 		KASSERT(ret, ("Unable to end access to grant references"));
1533 
1534 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1535 
1536 next:
1537 		if (m == NULL)
1538 			break;
1539 
1540 		m->m_len = rx->status;
1541 		m->m_data += rx->offset;
1542 		m0->m_pkthdr.len += rx->status;
1543 
1544 next_skip_queue:
1545 		if (!(rx->flags & NETRXF_more_data))
1546 			break;
1547 
1548 		if (*cons + frags == rp) {
1549 			if (net_ratelimit())
1550 				WPRINTK("Need more frags\n");
1551 			err = ENOENT;
1552 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1553 			       __func__, *cons, frags, rp);
1554 			break;
1555 		}
1556 		/*
1557 		 * Note that m can be NULL, if rx->status < 0 or if
1558 		 * rx->offset + rx->status > PAGE_SIZE above.
1559 		 */
1560 		m_prev = m;
1561 
1562 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1563 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1564 
1565 		/*
1566 		 * m_prev == NULL can happen if rx->status < 0 or if
1567 		 * rx->offset + * rx->status > PAGE_SIZE above.
1568 		 */
1569 		if (m_prev != NULL)
1570 			m_prev->m_next = m;
1571 
1572 		/*
1573 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1574 		 * rx->status > PAGE_SIZE above.
1575 		 */
1576 		if (m0 == NULL)
1577 			m0 = m;
1578 		m->m_next = NULL;
1579 		ref = xn_get_rx_ref(rxq, *cons + frags);
1580 		ref_cons = *cons + frags;
1581 		frags++;
1582 	}
1583 	*list = m0;
1584 	*cons += frags;
1585 
1586 	return (err);
1587 }
1588 
1589 /**
1590  * \brief Count the number of fragments in an mbuf chain.
1591  *
1592  * Surprisingly, there isn't an M* macro for this.
1593  */
1594 static inline int
1595 xn_count_frags(struct mbuf *m)
1596 {
1597 	int nfrags;
1598 
1599 	for (nfrags = 0; m != NULL; m = m->m_next)
1600 		nfrags++;
1601 
1602 	return (nfrags);
1603 }
1604 
1605 /**
1606  * Given an mbuf chain, make sure we have enough room and then push
1607  * it onto the transmit ring.
1608  */
1609 static int
1610 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1611 {
1612 	struct mbuf *m;
1613 	struct netfront_info *np = txq->info;
1614 	struct ifnet *ifp = np->xn_ifp;
1615 	u_int nfrags;
1616 	int otherend_id;
1617 
1618 	/**
1619 	 * Defragment the mbuf if necessary.
1620 	 */
1621 	nfrags = xn_count_frags(m_head);
1622 
1623 	/*
1624 	 * Check to see whether this request is longer than netback
1625 	 * can handle, and try to defrag it.
1626 	 */
1627 	/**
1628 	 * It is a bit lame, but the netback driver in Linux can't
1629 	 * deal with nfrags > MAX_TX_REQ_FRAGS, which is a quirk of
1630 	 * the Linux network stack.
1631 	 */
1632 	if (nfrags > np->maxfrags) {
1633 		m = m_defrag(m_head, M_NOWAIT);
1634 		if (!m) {
1635 			/*
1636 			 * Defrag failed, so free the mbuf and
1637 			 * therefore drop the packet.
1638 			 */
1639 			m_freem(m_head);
1640 			return (EMSGSIZE);
1641 		}
1642 		m_head = m;
1643 	}
1644 
1645 	/* Determine how many fragments now exist */
1646 	nfrags = xn_count_frags(m_head);
1647 
1648 	/*
1649 	 * Check to see whether the defragmented packet has too many
1650 	 * segments for the Linux netback driver.
1651 	 */
1652 	/**
1653 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1654 	 * of mbufs longer than Linux can handle.  Make sure we don't
1655 	 * pass a too-long chain over to the other side by dropping the
1656 	 * packet.  It doesn't look like there is currently a way to
1657 	 * tell the TCP stack to generate a shorter chain of packets.
1658 	 */
1659 	if (nfrags > MAX_TX_REQ_FRAGS) {
1660 #ifdef DEBUG
1661 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1662 		       "won't be able to handle it, dropping\n",
1663 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1664 #endif
1665 		m_freem(m_head);
1666 		return (EMSGSIZE);
1667 	}
1668 
1669 	/*
1670 	 * This check should be redundant.  We've already verified that we
1671 	 * have enough slots in the ring to handle a packet of maximum
1672 	 * size, and that our packet is less than the maximum size.  Keep
1673 	 * it in here as an assert for now just to make certain that
1674 	 * chain_cnt is accurate.
1675 	 */
1676 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1677 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1678 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1679                     (int) nfrags, (int) NET_TX_RING_SIZE));
1680 
1681 	/*
1682 	 * Start packing the mbufs in this chain into
1683 	 * the fragment pointers. Stop when we run out
1684 	 * of fragments or hit the end of the mbuf chain.
1685 	 */
1686 	m = m_head;
1687 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1688 	for (m = m_head; m; m = m->m_next) {
1689 		netif_tx_request_t *tx;
1690 		uintptr_t id;
1691 		grant_ref_t ref;
1692 		u_long mfn; /* XXX Wrong type? */
1693 
1694 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1695 		id = get_id_from_freelist(txq->mbufs);
1696 		if (id == 0)
1697 			panic("%s: was allocated the freelist head!\n",
1698 			    __func__);
1699 		txq->mbufs_cnt++;
1700 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1701 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1702 			    __func__);
1703 		txq->mbufs[id] = m;
1704 		tx->id = id;
1705 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1706 		KASSERT((short)ref >= 0, ("Negative ref"));
1707 		mfn = virt_to_mfn(mtod(m, vm_offset_t));
1708 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1709 		    mfn, GNTMAP_readonly);
1710 		tx->gref = txq->grant_ref[id] = ref;
1711 		tx->offset = mtod(m, vm_offset_t) & (PAGE_SIZE - 1);
1712 		tx->flags = 0;
1713 		if (m == m_head) {
1714 			/*
1715 			 * The first fragment has the entire packet
1716 			 * size, subsequent fragments have just the
1717 			 * fragment size. The backend works out the
1718 			 * true size of the first fragment by
1719 			 * subtracting the sizes of the other
1720 			 * fragments.
1721 			 */
1722 			tx->size = m->m_pkthdr.len;
1723 
1724 			/*
1725 			 * The first fragment contains the checksum flags
1726 			 * and is optionally followed by extra data for
1727 			 * TSO etc.
1728 			 */
1729 			/**
1730 			 * CSUM_TSO requires checksum offloading.
1731 			 * Some versions of FreeBSD fail to
1732 			 * set CSUM_TCP in the CSUM_TSO case,
1733 			 * so we have to test for CSUM_TSO
1734 			 * explicitly.
1735 			 */
1736 			if (m->m_pkthdr.csum_flags
1737 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1738 				tx->flags |= (NETTXF_csum_blank
1739 				    | NETTXF_data_validated);
1740 			}
1741 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1742 				struct netif_extra_info *gso =
1743 					(struct netif_extra_info *)
1744 					RING_GET_REQUEST(&txq->ring,
1745 							 ++txq->ring.req_prod_pvt);
1746 
1747 				tx->flags |= NETTXF_extra_info;
1748 
1749 				gso->u.gso.size = m->m_pkthdr.tso_segsz;
1750 				gso->u.gso.type =
1751 					XEN_NETIF_GSO_TYPE_TCPV4;
1752 				gso->u.gso.pad = 0;
1753 				gso->u.gso.features = 0;
1754 
1755 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1756 				gso->flags = 0;
1757 			}
1758 		} else {
1759 			tx->size = m->m_len;
1760 		}
1761 		if (m->m_next)
1762 			tx->flags |= NETTXF_more_data;
1763 
1764 		txq->ring.req_prod_pvt++;
1765 	}
1766 	BPF_MTAP(ifp, m_head);
1767 
1768 	xn_txeof(txq);
1769 
1770 	txq->stats.tx_bytes += m_head->m_pkthdr.len;
1771 	txq->stats.tx_packets++;
1772 
1773 	return (0);
1774 }
1775 
1776 /* equivalent of network_open() in Linux */
1777 static void
1778 xn_ifinit_locked(struct netfront_info *np)
1779 {
1780 	struct ifnet *ifp;
1781 	int i;
1782 	struct netfront_rxq *rxq;
1783 
1784 	XN_LOCK_ASSERT(np);
1785 
1786 	ifp = np->xn_ifp;
1787 
1788 	if (ifp->if_drv_flags & IFF_DRV_RUNNING)
1789 		return;
1790 
1791 	xn_stop(np);
1792 
1793 	for (i = 0; i < np->num_queues; i++) {
1794 		rxq = &np->rxq[i];
1795 		xn_alloc_rx_buffers(rxq);
1796 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1797 	}
1798 
1799 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
1800 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1801 	if_link_state_change(ifp, LINK_STATE_UP);
1802 }
1803 
1804 static void
1805 xn_ifinit(void *xsc)
1806 {
1807 	struct netfront_info *sc = xsc;
1808 
1809 	XN_LOCK(sc);
1810 	xn_ifinit_locked(sc);
1811 	XN_UNLOCK(sc);
1812 }
1813 
1814 static int
1815 xn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
1816 {
1817 	struct netfront_info *sc = ifp->if_softc;
1818 	struct ifreq *ifr = (struct ifreq *) data;
1819 #ifdef INET
1820 	struct ifaddr *ifa = (struct ifaddr *)data;
1821 #endif
1822 
1823 	int mask, error = 0;
1824 	switch(cmd) {
1825 	case SIOCSIFADDR:
1826 #ifdef INET
1827 		XN_LOCK(sc);
1828 		if (ifa->ifa_addr->sa_family == AF_INET) {
1829 			ifp->if_flags |= IFF_UP;
1830 			if (!(ifp->if_drv_flags & IFF_DRV_RUNNING))
1831 				xn_ifinit_locked(sc);
1832 			arp_ifinit(ifp, ifa);
1833 			XN_UNLOCK(sc);
1834 		} else {
1835 			XN_UNLOCK(sc);
1836 #endif
1837 			error = ether_ioctl(ifp, cmd, data);
1838 #ifdef INET
1839 		}
1840 #endif
1841 		break;
1842 	case SIOCSIFMTU:
1843 		ifp->if_mtu = ifr->ifr_mtu;
1844 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1845 		xn_ifinit(sc);
1846 		break;
1847 	case SIOCSIFFLAGS:
1848 		XN_LOCK(sc);
1849 		if (ifp->if_flags & IFF_UP) {
1850 			/*
1851 			 * If only the state of the PROMISC flag changed,
1852 			 * then just use the 'set promisc mode' command
1853 			 * instead of reinitializing the entire NIC. Doing
1854 			 * a full re-init means reloading the firmware and
1855 			 * waiting for it to start up, which may take a
1856 			 * second or two.
1857 			 */
1858 			xn_ifinit_locked(sc);
1859 		} else {
1860 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1861 				xn_stop(sc);
1862 			}
1863 		}
1864 		sc->xn_if_flags = ifp->if_flags;
1865 		XN_UNLOCK(sc);
1866 		error = 0;
1867 		break;
1868 	case SIOCSIFCAP:
1869 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
1870 		if (mask & IFCAP_TXCSUM) {
1871 			if (IFCAP_TXCSUM & ifp->if_capenable) {
1872 				ifp->if_capenable &= ~(IFCAP_TXCSUM|IFCAP_TSO4);
1873 				ifp->if_hwassist &= ~(CSUM_TCP | CSUM_UDP
1874 				    | CSUM_IP | CSUM_TSO);
1875 			} else {
1876 				ifp->if_capenable |= IFCAP_TXCSUM;
1877 				ifp->if_hwassist |= (CSUM_TCP | CSUM_UDP
1878 				    | CSUM_IP);
1879 			}
1880 		}
1881 		if (mask & IFCAP_RXCSUM) {
1882 			ifp->if_capenable ^= IFCAP_RXCSUM;
1883 		}
1884 		if (mask & IFCAP_TSO4) {
1885 			if (IFCAP_TSO4 & ifp->if_capenable) {
1886 				ifp->if_capenable &= ~IFCAP_TSO4;
1887 				ifp->if_hwassist &= ~CSUM_TSO;
1888 			} else if (IFCAP_TXCSUM & ifp->if_capenable) {
1889 				ifp->if_capenable |= IFCAP_TSO4;
1890 				ifp->if_hwassist |= CSUM_TSO;
1891 			} else {
1892 				IPRINTK("Xen requires tx checksum offload"
1893 				    " be enabled to use TSO\n");
1894 				error = EINVAL;
1895 			}
1896 		}
1897 		if (mask & IFCAP_LRO) {
1898 			ifp->if_capenable ^= IFCAP_LRO;
1899 
1900 		}
1901 		error = 0;
1902 		break;
1903 	case SIOCADDMULTI:
1904 	case SIOCDELMULTI:
1905 		break;
1906 	case SIOCSIFMEDIA:
1907 	case SIOCGIFMEDIA:
1908 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1909 		break;
1910 	default:
1911 		error = ether_ioctl(ifp, cmd, data);
1912 	}
1913 
1914 	return (error);
1915 }
1916 
1917 static void
1918 xn_stop(struct netfront_info *sc)
1919 {
1920 	struct ifnet *ifp;
1921 
1922 	XN_LOCK_ASSERT(sc);
1923 
1924 	ifp = sc->xn_ifp;
1925 
1926 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1927 	if_link_state_change(ifp, LINK_STATE_DOWN);
1928 }
1929 
1930 static void
1931 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1932 {
1933 	int requeue_idx, i;
1934 	grant_ref_t ref;
1935 	netif_rx_request_t *req;
1936 
1937 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1938 		struct mbuf *m;
1939 		u_long pfn;
1940 
1941 		if (rxq->mbufs[i] == NULL)
1942 			continue;
1943 
1944 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1945 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1946 
1947 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1948 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1949 
1950 		gnttab_grant_foreign_access_ref(ref,
1951 		    xenbus_get_otherend_id(rxq->info->xbdev),
1952 		    pfn, 0);
1953 
1954 		req->gref = ref;
1955 		req->id   = requeue_idx;
1956 
1957 		requeue_idx++;
1958 	}
1959 
1960 	rxq->ring.req_prod_pvt = requeue_idx;
1961 }
1962 
1963 /* START of Xenolinux helper functions adapted to FreeBSD */
1964 int
1965 xn_connect(struct netfront_info *np)
1966 {
1967 	int i, error;
1968 	u_int feature_rx_copy;
1969 	struct netfront_rxq *rxq;
1970 	struct netfront_txq *txq;
1971 
1972 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1973 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
1974 	if (error != 0)
1975 		feature_rx_copy = 0;
1976 
1977 	/* We only support rx copy. */
1978 	if (!feature_rx_copy)
1979 		return (EPROTONOSUPPORT);
1980 
1981 	/* Recovery procedure: */
1982 	error = talk_to_backend(np->xbdev, np);
1983 	if (error != 0)
1984 		return (error);
1985 
1986 	/* Step 1: Reinitialise variables. */
1987 	xn_query_features(np);
1988 	xn_configure_features(np);
1989 
1990 	/* Step 2: Release TX buffer */
1991 	for (i = 0; i < np->num_queues; i++) {
1992 		txq = &np->txq[i];
1993 		xn_release_tx_bufs(txq);
1994 	}
1995 
1996 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
1997 	for (i = 0; i < np->num_queues; i++) {
1998 		rxq = &np->rxq[i];
1999 		xn_rebuild_rx_bufs(rxq);
2000 	}
2001 
2002 	/* Step 4: All public and private state should now be sane.  Get
2003 	 * ready to start sending and receiving packets and give the driver
2004 	 * domain a kick because we've probably just requeued some
2005 	 * packets.
2006 	 */
2007 	netfront_carrier_on(np);
2008 	for (i = 0; i < np->num_queues; i++) {
2009 		txq = &np->txq[i];
2010 		xen_intr_signal(txq->xen_intr_handle);
2011 		XN_TX_LOCK(txq);
2012 		xn_txeof(txq);
2013 		XN_TX_UNLOCK(txq);
2014 		xn_alloc_rx_buffers(rxq);
2015 	}
2016 
2017 	return (0);
2018 }
2019 
2020 static void
2021 xn_query_features(struct netfront_info *np)
2022 {
2023 	int val;
2024 
2025 	device_printf(np->xbdev, "backend features:");
2026 
2027 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2028 		"feature-sg", NULL, "%d", &val) < 0)
2029 		val = 0;
2030 
2031 	np->maxfrags = 1;
2032 	if (val) {
2033 		np->maxfrags = MAX_TX_REQ_FRAGS;
2034 		printf(" feature-sg");
2035 	}
2036 
2037 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2038 		"feature-gso-tcpv4", NULL, "%d", &val) < 0)
2039 		val = 0;
2040 
2041 	np->xn_ifp->if_capabilities &= ~(IFCAP_TSO4|IFCAP_LRO);
2042 	if (val) {
2043 		np->xn_ifp->if_capabilities |= IFCAP_TSO4|IFCAP_LRO;
2044 		printf(" feature-gso-tcp4");
2045 	}
2046 
2047 	printf("\n");
2048 }
2049 
2050 static int
2051 xn_configure_features(struct netfront_info *np)
2052 {
2053 	int err, cap_enabled;
2054 #if (defined(INET) || defined(INET6))
2055 	int i;
2056 #endif
2057 
2058 	err = 0;
2059 
2060 	if (np->xn_resume &&
2061 	    ((np->xn_ifp->if_capenable & np->xn_ifp->if_capabilities)
2062 	    == np->xn_ifp->if_capenable)) {
2063 		/* Current options are available, no need to do anything. */
2064 		return (0);
2065 	}
2066 
2067 	/* Try to preserve as many options as possible. */
2068 	if (np->xn_resume)
2069 		cap_enabled = np->xn_ifp->if_capenable;
2070 	else
2071 		cap_enabled = UINT_MAX;
2072 
2073 #if (defined(INET) || defined(INET6))
2074 	for (i = 0; i < np->num_queues; i++)
2075 		if ((np->xn_ifp->if_capenable & IFCAP_LRO) ==
2076 		    (cap_enabled & IFCAP_LRO))
2077 			tcp_lro_free(&np->rxq[i].lro);
2078 #endif
2079     	np->xn_ifp->if_capenable =
2080 	    np->xn_ifp->if_capabilities & ~(IFCAP_LRO|IFCAP_TSO4) & cap_enabled;
2081 	np->xn_ifp->if_hwassist &= ~CSUM_TSO;
2082 #if (defined(INET) || defined(INET6))
2083 	for (i = 0; i < np->num_queues; i++) {
2084 		if (xn_enable_lro && (np->xn_ifp->if_capabilities & IFCAP_LRO) ==
2085 		    (cap_enabled & IFCAP_LRO)) {
2086 			err = tcp_lro_init(&np->rxq[i].lro);
2087 			if (err != 0) {
2088 				device_printf(np->xbdev, "LRO initialization failed\n");
2089 			} else {
2090 				np->rxq[i].lro.ifp = np->xn_ifp;
2091 				np->xn_ifp->if_capenable |= IFCAP_LRO;
2092 			}
2093 		}
2094 	}
2095 	if ((np->xn_ifp->if_capabilities & IFCAP_TSO4) ==
2096 	    (cap_enabled & IFCAP_TSO4)) {
2097 		np->xn_ifp->if_capenable |= IFCAP_TSO4;
2098 		np->xn_ifp->if_hwassist |= CSUM_TSO;
2099 	}
2100 #endif
2101 	return (err);
2102 }
2103 
2104 static int
2105 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2106 {
2107 	struct netfront_info *np;
2108 	struct ifnet *ifp;
2109 	struct buf_ring *br;
2110 	int error, notify;
2111 
2112 	np = txq->info;
2113 	br = txq->br;
2114 	ifp = np->xn_ifp;
2115 	error = 0;
2116 
2117 	XN_TX_LOCK_ASSERT(txq);
2118 
2119 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
2120 	    !netfront_carrier_ok(np)) {
2121 		if (m != NULL)
2122 			error = drbr_enqueue(ifp, br, m);
2123 		return (error);
2124 	}
2125 
2126 	if (m != NULL) {
2127 		error = drbr_enqueue(ifp, br, m);
2128 		if (error != 0)
2129 			return (error);
2130 	}
2131 
2132 	while ((m = drbr_peek(ifp, br)) != NULL) {
2133 		if (!xn_tx_slot_available(txq)) {
2134 			drbr_putback(ifp, br, m);
2135 			break;
2136 		}
2137 
2138 		error = xn_assemble_tx_request(txq, m);
2139 		/* xn_assemble_tx_request always consumes the mbuf*/
2140 		if (error != 0) {
2141 			drbr_advance(ifp, br);
2142 			break;
2143 		}
2144 
2145 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2146 		if (notify)
2147 			xen_intr_signal(txq->xen_intr_handle);
2148 
2149 		drbr_advance(ifp, br);
2150 	}
2151 
2152 	if (RING_FULL(&txq->ring))
2153 		txq->full = true;
2154 
2155 	return (0);
2156 }
2157 
2158 static int
2159 xn_txq_mq_start(struct ifnet *ifp, struct mbuf *m)
2160 {
2161 	struct netfront_info *np;
2162 	struct netfront_txq *txq;
2163 	int i, npairs, error;
2164 
2165 	np = ifp->if_softc;
2166 	npairs = np->num_queues;
2167 
2168 	/* check if flowid is set */
2169 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2170 		i = m->m_pkthdr.flowid % npairs;
2171 	else
2172 		i = curcpu % npairs;
2173 
2174 	txq = &np->txq[i];
2175 
2176 	if (XN_TX_TRYLOCK(txq) != 0) {
2177 		error = xn_txq_mq_start_locked(txq, m);
2178 		XN_TX_UNLOCK(txq);
2179 	} else {
2180 		error = drbr_enqueue(ifp, txq->br, m);
2181 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2182 	}
2183 
2184 	return (error);
2185 }
2186 
2187 static void
2188 xn_qflush(struct ifnet *ifp)
2189 {
2190 	struct netfront_info *np;
2191 	struct netfront_txq *txq;
2192 	struct mbuf *m;
2193 	int i;
2194 
2195 	np = ifp->if_softc;
2196 
2197 	for (i = 0; i < np->num_queues; i++) {
2198 		txq = &np->txq[i];
2199 
2200 		XN_TX_LOCK(txq);
2201 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2202 			m_freem(m);
2203 		XN_TX_UNLOCK(txq);
2204 	}
2205 
2206 	if_qflush(ifp);
2207 }
2208 
2209 /**
2210  * Create a network device.
2211  * @param dev  Newbus device representing this virtual NIC.
2212  */
2213 int
2214 create_netdev(device_t dev)
2215 {
2216 	struct netfront_info *np;
2217 	int err;
2218 	struct ifnet *ifp;
2219 
2220 	np = device_get_softc(dev);
2221 
2222 	np->xbdev         = dev;
2223 
2224 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2225 
2226 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2227 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2228 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2229 
2230 	np->rx_min_target = RX_MIN_TARGET;
2231 	np->rx_max_target = RX_MAX_TARGET;
2232 
2233 	err = xen_net_read_mac(dev, np->mac);
2234 	if (err != 0)
2235 		goto error;
2236 
2237 	/* Set up ifnet structure */
2238 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2239     	ifp->if_softc = np;
2240     	if_initname(ifp, "xn",  device_get_unit(dev));
2241     	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2242     	ifp->if_ioctl = xn_ioctl;
2243 
2244 	ifp->if_transmit = xn_txq_mq_start;
2245 	ifp->if_qflush = xn_qflush;
2246 
2247     	ifp->if_init = xn_ifinit;
2248 
2249     	ifp->if_hwassist = XN_CSUM_FEATURES;
2250     	ifp->if_capabilities = IFCAP_HWCSUM;
2251 	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2252 	ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
2253 	ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
2254 
2255     	ether_ifattach(ifp, np->mac);
2256 	netfront_carrier_off(np);
2257 
2258 	return (0);
2259 
2260 error:
2261 	KASSERT(err != 0, ("Error path with no error code specified"));
2262 	return (err);
2263 }
2264 
2265 static int
2266 netfront_detach(device_t dev)
2267 {
2268 	struct netfront_info *info = device_get_softc(dev);
2269 
2270 	DPRINTK("%s\n", xenbus_get_node(dev));
2271 
2272 	netif_free(info);
2273 
2274 	return 0;
2275 }
2276 
2277 static void
2278 netif_free(struct netfront_info *np)
2279 {
2280 
2281 	XN_LOCK(np);
2282 	xn_stop(np);
2283 	XN_UNLOCK(np);
2284 	netif_disconnect_backend(np);
2285 	free(np->rxq, M_DEVBUF);
2286 	free(np->txq, M_DEVBUF);
2287 	if (np->xn_ifp != NULL) {
2288 		ether_ifdetach(np->xn_ifp);
2289 		if_free(np->xn_ifp);
2290 		np->xn_ifp = NULL;
2291 	}
2292 	ifmedia_removeall(&np->sc_media);
2293 }
2294 
2295 static void
2296 netif_disconnect_backend(struct netfront_info *np)
2297 {
2298 	u_int i;
2299 
2300 	for (i = 0; i < np->num_queues; i++) {
2301 		XN_RX_LOCK(&np->rxq[i]);
2302 		XN_TX_LOCK(&np->txq[i]);
2303 	}
2304 	netfront_carrier_off(np);
2305 	for (i = 0; i < np->num_queues; i++) {
2306 		XN_RX_UNLOCK(&np->rxq[i]);
2307 		XN_TX_UNLOCK(&np->txq[i]);
2308 	}
2309 
2310 	for (i = 0; i < np->num_queues; i++) {
2311 		disconnect_rxq(&np->rxq[i]);
2312 		disconnect_txq(&np->txq[i]);
2313 	}
2314 }
2315 
2316 static int
2317 xn_ifmedia_upd(struct ifnet *ifp)
2318 {
2319 
2320 	return (0);
2321 }
2322 
2323 static void
2324 xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
2325 {
2326 
2327 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2328 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2329 }
2330 
2331 /* ** Driver registration ** */
2332 static device_method_t netfront_methods[] = {
2333 	/* Device interface */
2334 	DEVMETHOD(device_probe,         netfront_probe),
2335 	DEVMETHOD(device_attach,        netfront_attach),
2336 	DEVMETHOD(device_detach,        netfront_detach),
2337 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2338 	DEVMETHOD(device_suspend,       netfront_suspend),
2339 	DEVMETHOD(device_resume,        netfront_resume),
2340 
2341 	/* Xenbus interface */
2342 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2343 
2344 	DEVMETHOD_END
2345 };
2346 
2347 static driver_t netfront_driver = {
2348 	"xn",
2349 	netfront_methods,
2350 	sizeof(struct netfront_info),
2351 };
2352 devclass_t netfront_devclass;
2353 
2354 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, netfront_devclass, NULL,
2355     NULL);
2356