xref: /freebsd/sys/dev/xen/netfront/netfront.c (revision f39bffc62c1395bde25d152c7f68fdf7cbaab414)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2004-2006 Kip Macy
5  * Copyright (c) 2015 Wei Liu <wei.liu2@citrix.com>
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_inet.h"
34 #include "opt_inet6.h"
35 
36 #include <sys/param.h>
37 #include <sys/sockio.h>
38 #include <sys/limits.h>
39 #include <sys/mbuf.h>
40 #include <sys/malloc.h>
41 #include <sys/module.h>
42 #include <sys/kernel.h>
43 #include <sys/socket.h>
44 #include <sys/sysctl.h>
45 #include <sys/taskqueue.h>
46 
47 #include <net/if.h>
48 #include <net/if_var.h>
49 #include <net/if_arp.h>
50 #include <net/ethernet.h>
51 #include <net/if_media.h>
52 #include <net/bpf.h>
53 #include <net/if_types.h>
54 
55 #include <netinet/in.h>
56 #include <netinet/ip.h>
57 #include <netinet/if_ether.h>
58 #include <netinet/tcp.h>
59 #include <netinet/tcp_lro.h>
60 
61 #include <vm/vm.h>
62 #include <vm/pmap.h>
63 
64 #include <sys/bus.h>
65 
66 #include <xen/xen-os.h>
67 #include <xen/hypervisor.h>
68 #include <xen/xen_intr.h>
69 #include <xen/gnttab.h>
70 #include <xen/interface/memory.h>
71 #include <xen/interface/io/netif.h>
72 #include <xen/xenbus/xenbusvar.h>
73 
74 #include "xenbus_if.h"
75 
76 /* Features supported by all backends.  TSO and LRO can be negotiated */
77 #define XN_CSUM_FEATURES	(CSUM_TCP | CSUM_UDP)
78 
79 #define NET_TX_RING_SIZE __CONST_RING_SIZE(netif_tx, PAGE_SIZE)
80 #define NET_RX_RING_SIZE __CONST_RING_SIZE(netif_rx, PAGE_SIZE)
81 
82 #define NET_RX_SLOTS_MIN (XEN_NETIF_NR_SLOTS_MIN + 1)
83 
84 /*
85  * Should the driver do LRO on the RX end
86  *  this can be toggled on the fly, but the
87  *  interface must be reset (down/up) for it
88  *  to take effect.
89  */
90 static int xn_enable_lro = 1;
91 TUNABLE_INT("hw.xn.enable_lro", &xn_enable_lro);
92 
93 /*
94  * Number of pairs of queues.
95  */
96 static unsigned long xn_num_queues = 4;
97 TUNABLE_ULONG("hw.xn.num_queues", &xn_num_queues);
98 
99 /**
100  * \brief The maximum allowed data fragments in a single transmit
101  *        request.
102  *
103  * This limit is imposed by the backend driver.  We assume here that
104  * we are dealing with a Linux driver domain and have set our limit
105  * to mirror the Linux MAX_SKB_FRAGS constant.
106  */
107 #define	MAX_TX_REQ_FRAGS (65536 / PAGE_SIZE + 2)
108 
109 #define RX_COPY_THRESHOLD 256
110 
111 #define net_ratelimit() 0
112 
113 struct netfront_rxq;
114 struct netfront_txq;
115 struct netfront_info;
116 struct netfront_rx_info;
117 
118 static void xn_txeof(struct netfront_txq *);
119 static void xn_rxeof(struct netfront_rxq *);
120 static void xn_alloc_rx_buffers(struct netfront_rxq *);
121 static void xn_alloc_rx_buffers_callout(void *arg);
122 
123 static void xn_release_rx_bufs(struct netfront_rxq *);
124 static void xn_release_tx_bufs(struct netfront_txq *);
125 
126 static void xn_rxq_intr(struct netfront_rxq *);
127 static void xn_txq_intr(struct netfront_txq *);
128 static void xn_intr(void *);
129 static inline int xn_count_frags(struct mbuf *m);
130 static int xn_assemble_tx_request(struct netfront_txq *, struct mbuf *);
131 static int xn_ioctl(struct ifnet *, u_long, caddr_t);
132 static void xn_ifinit_locked(struct netfront_info *);
133 static void xn_ifinit(void *);
134 static void xn_stop(struct netfront_info *);
135 static void xn_query_features(struct netfront_info *np);
136 static int xn_configure_features(struct netfront_info *np);
137 static void netif_free(struct netfront_info *info);
138 static int netfront_detach(device_t dev);
139 
140 static int xn_txq_mq_start_locked(struct netfront_txq *, struct mbuf *);
141 static int xn_txq_mq_start(struct ifnet *, struct mbuf *);
142 
143 static int talk_to_backend(device_t dev, struct netfront_info *info);
144 static int create_netdev(device_t dev);
145 static void netif_disconnect_backend(struct netfront_info *info);
146 static int setup_device(device_t dev, struct netfront_info *info,
147     unsigned long);
148 static int xn_ifmedia_upd(struct ifnet *ifp);
149 static void xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr);
150 
151 static int xn_connect(struct netfront_info *);
152 static void xn_kick_rings(struct netfront_info *);
153 
154 static int xn_get_responses(struct netfront_rxq *,
155     struct netfront_rx_info *, RING_IDX, RING_IDX *,
156     struct mbuf **);
157 
158 #define virt_to_mfn(x) (vtophys(x) >> PAGE_SHIFT)
159 
160 #define INVALID_P2M_ENTRY (~0UL)
161 #define XN_QUEUE_NAME_LEN  8	/* xn{t,r}x_%u, allow for two digits */
162 struct netfront_rxq {
163 	struct netfront_info 	*info;
164 	u_int			id;
165 	char			name[XN_QUEUE_NAME_LEN];
166 	struct mtx		lock;
167 
168 	int			ring_ref;
169 	netif_rx_front_ring_t 	ring;
170 	xen_intr_handle_t	xen_intr_handle;
171 
172 	grant_ref_t 		gref_head;
173 	grant_ref_t 		grant_ref[NET_RX_RING_SIZE + 1];
174 
175 	struct mbuf		*mbufs[NET_RX_RING_SIZE + 1];
176 
177 	struct lro_ctrl		lro;
178 
179 	struct callout		rx_refill;
180 };
181 
182 struct netfront_txq {
183 	struct netfront_info 	*info;
184 	u_int 			id;
185 	char			name[XN_QUEUE_NAME_LEN];
186 	struct mtx		lock;
187 
188 	int			ring_ref;
189 	netif_tx_front_ring_t	ring;
190 	xen_intr_handle_t 	xen_intr_handle;
191 
192 	grant_ref_t		gref_head;
193 	grant_ref_t		grant_ref[NET_TX_RING_SIZE + 1];
194 
195 	struct mbuf		*mbufs[NET_TX_RING_SIZE + 1];
196 	int			mbufs_cnt;
197 	struct buf_ring		*br;
198 
199 	struct taskqueue 	*tq;
200 	struct task       	defrtask;
201 
202 	bool			full;
203 };
204 
205 struct netfront_info {
206 	struct ifnet 		*xn_ifp;
207 
208 	struct mtx   		sc_lock;
209 
210 	u_int  num_queues;
211 	struct netfront_rxq 	*rxq;
212 	struct netfront_txq 	*txq;
213 
214 	u_int			carrier;
215 	u_int			maxfrags;
216 
217 	device_t		xbdev;
218 	uint8_t			mac[ETHER_ADDR_LEN];
219 
220 	int			xn_if_flags;
221 
222 	struct ifmedia		sc_media;
223 
224 	bool			xn_reset;
225 };
226 
227 struct netfront_rx_info {
228 	struct netif_rx_response rx;
229 	struct netif_extra_info extras[XEN_NETIF_EXTRA_TYPE_MAX - 1];
230 };
231 
232 #define XN_RX_LOCK(_q)         mtx_lock(&(_q)->lock)
233 #define XN_RX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
234 
235 #define XN_TX_LOCK(_q)         mtx_lock(&(_q)->lock)
236 #define XN_TX_TRYLOCK(_q)      mtx_trylock(&(_q)->lock)
237 #define XN_TX_UNLOCK(_q)       mtx_unlock(&(_q)->lock)
238 
239 #define XN_LOCK(_sc)           mtx_lock(&(_sc)->sc_lock);
240 #define XN_UNLOCK(_sc)         mtx_unlock(&(_sc)->sc_lock);
241 
242 #define XN_LOCK_ASSERT(_sc)    mtx_assert(&(_sc)->sc_lock, MA_OWNED);
243 #define XN_RX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
244 #define XN_TX_LOCK_ASSERT(_q)  mtx_assert(&(_q)->lock, MA_OWNED);
245 
246 #define netfront_carrier_on(netif)	((netif)->carrier = 1)
247 #define netfront_carrier_off(netif)	((netif)->carrier = 0)
248 #define netfront_carrier_ok(netif)	((netif)->carrier)
249 
250 /* Access macros for acquiring freeing slots in xn_free_{tx,rx}_idxs[]. */
251 
252 static inline void
253 add_id_to_freelist(struct mbuf **list, uintptr_t id)
254 {
255 
256 	KASSERT(id != 0,
257 		("%s: the head item (0) must always be free.", __func__));
258 	list[id] = list[0];
259 	list[0]  = (struct mbuf *)id;
260 }
261 
262 static inline unsigned short
263 get_id_from_freelist(struct mbuf **list)
264 {
265 	uintptr_t id;
266 
267 	id = (uintptr_t)list[0];
268 	KASSERT(id != 0,
269 		("%s: the head item (0) must always remain free.", __func__));
270 	list[0] = list[id];
271 	return (id);
272 }
273 
274 static inline int
275 xn_rxidx(RING_IDX idx)
276 {
277 
278 	return idx & (NET_RX_RING_SIZE - 1);
279 }
280 
281 static inline struct mbuf *
282 xn_get_rx_mbuf(struct netfront_rxq *rxq, RING_IDX ri)
283 {
284 	int i;
285 	struct mbuf *m;
286 
287 	i = xn_rxidx(ri);
288 	m = rxq->mbufs[i];
289 	rxq->mbufs[i] = NULL;
290 	return (m);
291 }
292 
293 static inline grant_ref_t
294 xn_get_rx_ref(struct netfront_rxq *rxq, RING_IDX ri)
295 {
296 	int i = xn_rxidx(ri);
297 	grant_ref_t ref = rxq->grant_ref[i];
298 
299 	KASSERT(ref != GRANT_REF_INVALID, ("Invalid grant reference!\n"));
300 	rxq->grant_ref[i] = GRANT_REF_INVALID;
301 	return (ref);
302 }
303 
304 #define IPRINTK(fmt, args...) \
305     printf("[XEN] " fmt, ##args)
306 #ifdef INVARIANTS
307 #define WPRINTK(fmt, args...) \
308     printf("[XEN] " fmt, ##args)
309 #else
310 #define WPRINTK(fmt, args...)
311 #endif
312 #ifdef DEBUG
313 #define DPRINTK(fmt, args...) \
314     printf("[XEN] %s: " fmt, __func__, ##args)
315 #else
316 #define DPRINTK(fmt, args...)
317 #endif
318 
319 /**
320  * Read the 'mac' node at the given device's node in the store, and parse that
321  * as colon-separated octets, placing result the given mac array.  mac must be
322  * a preallocated array of length ETH_ALEN (as declared in linux/if_ether.h).
323  * Return 0 on success, or errno on error.
324  */
325 static int
326 xen_net_read_mac(device_t dev, uint8_t mac[])
327 {
328 	int error, i;
329 	char *s, *e, *macstr;
330 	const char *path;
331 
332 	path = xenbus_get_node(dev);
333 	error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
334 	if (error == ENOENT) {
335 		/*
336 		 * Deal with missing mac XenStore nodes on devices with
337 		 * HVM emulation (the 'ioemu' configuration attribute)
338 		 * enabled.
339 		 *
340 		 * The HVM emulator may execute in a stub device model
341 		 * domain which lacks the permission, only given to Dom0,
342 		 * to update the guest's XenStore tree.  For this reason,
343 		 * the HVM emulator doesn't even attempt to write the
344 		 * front-side mac node, even when operating in Dom0.
345 		 * However, there should always be a mac listed in the
346 		 * backend tree.  Fallback to this version if our query
347 		 * of the front side XenStore location doesn't find
348 		 * anything.
349 		 */
350 		path = xenbus_get_otherend_path(dev);
351 		error = xs_read(XST_NIL, path, "mac", NULL, (void **) &macstr);
352 	}
353 	if (error != 0) {
354 		xenbus_dev_fatal(dev, error, "parsing %s/mac", path);
355 		return (error);
356 	}
357 
358 	s = macstr;
359 	for (i = 0; i < ETHER_ADDR_LEN; i++) {
360 		mac[i] = strtoul(s, &e, 16);
361 		if (s == e || (e[0] != ':' && e[0] != 0)) {
362 			free(macstr, M_XENBUS);
363 			return (ENOENT);
364 		}
365 		s = &e[1];
366 	}
367 	free(macstr, M_XENBUS);
368 	return (0);
369 }
370 
371 /**
372  * Entry point to this code when a new device is created.  Allocate the basic
373  * structures and the ring buffers for communication with the backend, and
374  * inform the backend of the appropriate details for those.  Switch to
375  * Connected state.
376  */
377 static int
378 netfront_probe(device_t dev)
379 {
380 
381 	if (xen_hvm_domain() && xen_disable_pv_nics != 0)
382 		return (ENXIO);
383 
384 	if (!strcmp(xenbus_get_type(dev), "vif")) {
385 		device_set_desc(dev, "Virtual Network Interface");
386 		return (0);
387 	}
388 
389 	return (ENXIO);
390 }
391 
392 static int
393 netfront_attach(device_t dev)
394 {
395 	int err;
396 
397 	err = create_netdev(dev);
398 	if (err != 0) {
399 		xenbus_dev_fatal(dev, err, "creating netdev");
400 		return (err);
401 	}
402 
403 	SYSCTL_ADD_INT(device_get_sysctl_ctx(dev),
404 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
405 	    OID_AUTO, "enable_lro", CTLFLAG_RW,
406 	    &xn_enable_lro, 0, "Large Receive Offload");
407 
408 	SYSCTL_ADD_ULONG(device_get_sysctl_ctx(dev),
409 	    SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
410 	    OID_AUTO, "num_queues", CTLFLAG_RD,
411 	    &xn_num_queues, "Number of pairs of queues");
412 
413 	return (0);
414 }
415 
416 static int
417 netfront_suspend(device_t dev)
418 {
419 	struct netfront_info *np = device_get_softc(dev);
420 	u_int i;
421 
422 	for (i = 0; i < np->num_queues; i++) {
423 		XN_RX_LOCK(&np->rxq[i]);
424 		XN_TX_LOCK(&np->txq[i]);
425 	}
426 	netfront_carrier_off(np);
427 	for (i = 0; i < np->num_queues; i++) {
428 		XN_RX_UNLOCK(&np->rxq[i]);
429 		XN_TX_UNLOCK(&np->txq[i]);
430 	}
431 	return (0);
432 }
433 
434 /**
435  * We are reconnecting to the backend, due to a suspend/resume, or a backend
436  * driver restart.  We tear down our netif structure and recreate it, but
437  * leave the device-layer structures intact so that this is transparent to the
438  * rest of the kernel.
439  */
440 static int
441 netfront_resume(device_t dev)
442 {
443 	struct netfront_info *info = device_get_softc(dev);
444 	u_int i;
445 
446 	if (xen_suspend_cancelled) {
447 		for (i = 0; i < info->num_queues; i++) {
448 			XN_RX_LOCK(&info->rxq[i]);
449 			XN_TX_LOCK(&info->txq[i]);
450 		}
451 		netfront_carrier_on(info);
452 		for (i = 0; i < info->num_queues; i++) {
453 			XN_RX_UNLOCK(&info->rxq[i]);
454 			XN_TX_UNLOCK(&info->txq[i]);
455 		}
456 		return (0);
457 	}
458 
459 	netif_disconnect_backend(info);
460 	return (0);
461 }
462 
463 static int
464 write_queue_xenstore_keys(device_t dev,
465     struct netfront_rxq *rxq,
466     struct netfront_txq *txq,
467     struct xs_transaction *xst, bool hierarchy)
468 {
469 	int err;
470 	const char *message;
471 	const char *node = xenbus_get_node(dev);
472 	char *path;
473 	size_t path_size;
474 
475 	KASSERT(rxq->id == txq->id, ("Mismatch between RX and TX queue ids"));
476 	/* Split event channel support is not yet there. */
477 	KASSERT(rxq->xen_intr_handle == txq->xen_intr_handle,
478 	    ("Split event channels are not supported"));
479 
480 	if (hierarchy) {
481 		path_size = strlen(node) + 10;
482 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
483 		snprintf(path, path_size, "%s/queue-%u", node, rxq->id);
484 	} else {
485 		path_size = strlen(node) + 1;
486 		path = malloc(path_size, M_DEVBUF, M_WAITOK|M_ZERO);
487 		snprintf(path, path_size, "%s", node);
488 	}
489 
490 	err = xs_printf(*xst, path, "tx-ring-ref","%u", txq->ring_ref);
491 	if (err != 0) {
492 		message = "writing tx ring-ref";
493 		goto error;
494 	}
495 	err = xs_printf(*xst, path, "rx-ring-ref","%u", rxq->ring_ref);
496 	if (err != 0) {
497 		message = "writing rx ring-ref";
498 		goto error;
499 	}
500 	err = xs_printf(*xst, path, "event-channel", "%u",
501 	    xen_intr_port(rxq->xen_intr_handle));
502 	if (err != 0) {
503 		message = "writing event-channel";
504 		goto error;
505 	}
506 
507 	free(path, M_DEVBUF);
508 
509 	return (0);
510 
511 error:
512 	free(path, M_DEVBUF);
513 	xenbus_dev_fatal(dev, err, "%s", message);
514 
515 	return (err);
516 }
517 
518 /* Common code used when first setting up, and when resuming. */
519 static int
520 talk_to_backend(device_t dev, struct netfront_info *info)
521 {
522 	const char *message;
523 	struct xs_transaction xst;
524 	const char *node = xenbus_get_node(dev);
525 	int err;
526 	unsigned long num_queues, max_queues = 0;
527 	unsigned int i;
528 
529 	err = xen_net_read_mac(dev, info->mac);
530 	if (err != 0) {
531 		xenbus_dev_fatal(dev, err, "parsing %s/mac", node);
532 		goto out;
533 	}
534 
535 	err = xs_scanf(XST_NIL, xenbus_get_otherend_path(info->xbdev),
536 	    "multi-queue-max-queues", NULL, "%lu", &max_queues);
537 	if (err != 0)
538 		max_queues = 1;
539 	num_queues = xn_num_queues;
540 	if (num_queues > max_queues)
541 		num_queues = max_queues;
542 
543 	err = setup_device(dev, info, num_queues);
544 	if (err != 0)
545 		goto out;
546 
547  again:
548 	err = xs_transaction_start(&xst);
549 	if (err != 0) {
550 		xenbus_dev_fatal(dev, err, "starting transaction");
551 		goto free;
552 	}
553 
554 	if (info->num_queues == 1) {
555 		err = write_queue_xenstore_keys(dev, &info->rxq[0],
556 		    &info->txq[0], &xst, false);
557 		if (err != 0)
558 			goto abort_transaction_no_def_error;
559 	} else {
560 		err = xs_printf(xst, node, "multi-queue-num-queues",
561 		    "%u", info->num_queues);
562 		if (err != 0) {
563 			message = "writing multi-queue-num-queues";
564 			goto abort_transaction;
565 		}
566 
567 		for (i = 0; i < info->num_queues; i++) {
568 			err = write_queue_xenstore_keys(dev, &info->rxq[i],
569 			    &info->txq[i], &xst, true);
570 			if (err != 0)
571 				goto abort_transaction_no_def_error;
572 		}
573 	}
574 
575 	err = xs_printf(xst, node, "request-rx-copy", "%u", 1);
576 	if (err != 0) {
577 		message = "writing request-rx-copy";
578 		goto abort_transaction;
579 	}
580 	err = xs_printf(xst, node, "feature-rx-notify", "%d", 1);
581 	if (err != 0) {
582 		message = "writing feature-rx-notify";
583 		goto abort_transaction;
584 	}
585 	err = xs_printf(xst, node, "feature-sg", "%d", 1);
586 	if (err != 0) {
587 		message = "writing feature-sg";
588 		goto abort_transaction;
589 	}
590 	if ((info->xn_ifp->if_capenable & IFCAP_LRO) != 0) {
591 		err = xs_printf(xst, node, "feature-gso-tcpv4", "%d", 1);
592 		if (err != 0) {
593 			message = "writing feature-gso-tcpv4";
594 			goto abort_transaction;
595 		}
596 	}
597 	if ((info->xn_ifp->if_capenable & IFCAP_RXCSUM) == 0) {
598 		err = xs_printf(xst, node, "feature-no-csum-offload", "%d", 1);
599 		if (err != 0) {
600 			message = "writing feature-no-csum-offload";
601 			goto abort_transaction;
602 		}
603 	}
604 
605 	err = xs_transaction_end(xst, 0);
606 	if (err != 0) {
607 		if (err == EAGAIN)
608 			goto again;
609 		xenbus_dev_fatal(dev, err, "completing transaction");
610 		goto free;
611 	}
612 
613 	return 0;
614 
615  abort_transaction:
616 	xenbus_dev_fatal(dev, err, "%s", message);
617  abort_transaction_no_def_error:
618 	xs_transaction_end(xst, 1);
619  free:
620 	netif_free(info);
621  out:
622 	return (err);
623 }
624 
625 static void
626 xn_rxq_intr(struct netfront_rxq *rxq)
627 {
628 
629 	XN_RX_LOCK(rxq);
630 	xn_rxeof(rxq);
631 	XN_RX_UNLOCK(rxq);
632 }
633 
634 static void
635 xn_txq_start(struct netfront_txq *txq)
636 {
637 	struct netfront_info *np = txq->info;
638 	struct ifnet *ifp = np->xn_ifp;
639 
640 	XN_TX_LOCK_ASSERT(txq);
641 	if (!drbr_empty(ifp, txq->br))
642 		xn_txq_mq_start_locked(txq, NULL);
643 }
644 
645 static void
646 xn_txq_intr(struct netfront_txq *txq)
647 {
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(rxq->ring_ref, NULL);
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 }
688 
689 static void
690 destroy_rxqs(struct netfront_info *np)
691 {
692 	int i;
693 
694 	for (i = 0; i < np->num_queues; i++)
695 		destroy_rxq(&np->rxq[i]);
696 
697 	free(np->rxq, M_DEVBUF);
698 	np->rxq = NULL;
699 }
700 
701 static int
702 setup_rxqs(device_t dev, struct netfront_info *info,
703 	   unsigned long num_queues)
704 {
705 	int q, i;
706 	int error;
707 	netif_rx_sring_t *rxs;
708 	struct netfront_rxq *rxq;
709 
710 	info->rxq = malloc(sizeof(struct netfront_rxq) * num_queues,
711 	    M_DEVBUF, M_WAITOK|M_ZERO);
712 
713 	for (q = 0; q < num_queues; q++) {
714 		rxq = &info->rxq[q];
715 
716 		rxq->id = q;
717 		rxq->info = info;
718 		rxq->ring_ref = GRANT_REF_INVALID;
719 		rxq->ring.sring = NULL;
720 		snprintf(rxq->name, XN_QUEUE_NAME_LEN, "xnrx_%u", q);
721 		mtx_init(&rxq->lock, rxq->name, "netfront receive lock",
722 		    MTX_DEF);
723 
724 		for (i = 0; i <= NET_RX_RING_SIZE; i++) {
725 			rxq->mbufs[i] = NULL;
726 			rxq->grant_ref[i] = GRANT_REF_INVALID;
727 		}
728 
729 		/* Start resources allocation */
730 
731 		if (gnttab_alloc_grant_references(NET_RX_RING_SIZE,
732 		    &rxq->gref_head) != 0) {
733 			device_printf(dev, "allocating rx gref");
734 			error = ENOMEM;
735 			goto fail;
736 		}
737 
738 		rxs = (netif_rx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
739 		    M_WAITOK|M_ZERO);
740 		SHARED_RING_INIT(rxs);
741 		FRONT_RING_INIT(&rxq->ring, rxs, PAGE_SIZE);
742 
743 		error = xenbus_grant_ring(dev, virt_to_mfn(rxs),
744 		    &rxq->ring_ref);
745 		if (error != 0) {
746 			device_printf(dev, "granting rx ring page");
747 			goto fail_grant_ring;
748 		}
749 
750 		callout_init(&rxq->rx_refill, 1);
751 	}
752 
753 	return (0);
754 
755 fail_grant_ring:
756 	gnttab_free_grant_references(rxq->gref_head);
757 	free(rxq->ring.sring, M_DEVBUF);
758 fail:
759 	for (; q >= 0; q--) {
760 		disconnect_rxq(&info->rxq[q]);
761 		destroy_rxq(&info->rxq[q]);
762 	}
763 
764 	free(info->rxq, M_DEVBUF);
765 	return (error);
766 }
767 
768 static void
769 disconnect_txq(struct netfront_txq *txq)
770 {
771 
772 	xn_release_tx_bufs(txq);
773 	gnttab_free_grant_references(txq->gref_head);
774 	gnttab_end_foreign_access(txq->ring_ref, NULL);
775 	xen_intr_unbind(&txq->xen_intr_handle);
776 }
777 
778 static void
779 destroy_txq(struct netfront_txq *txq)
780 {
781 
782 	free(txq->ring.sring, M_DEVBUF);
783 	buf_ring_free(txq->br, M_DEVBUF);
784 	taskqueue_drain_all(txq->tq);
785 	taskqueue_free(txq->tq);
786 }
787 
788 static void
789 destroy_txqs(struct netfront_info *np)
790 {
791 	int i;
792 
793 	for (i = 0; i < np->num_queues; i++)
794 		destroy_txq(&np->txq[i]);
795 
796 	free(np->txq, M_DEVBUF);
797 	np->txq = NULL;
798 }
799 
800 static int
801 setup_txqs(device_t dev, struct netfront_info *info,
802 	   unsigned long num_queues)
803 {
804 	int q, i;
805 	int error;
806 	netif_tx_sring_t *txs;
807 	struct netfront_txq *txq;
808 
809 	info->txq = malloc(sizeof(struct netfront_txq) * num_queues,
810 	    M_DEVBUF, M_WAITOK|M_ZERO);
811 
812 	for (q = 0; q < num_queues; q++) {
813 		txq = &info->txq[q];
814 
815 		txq->id = q;
816 		txq->info = info;
817 
818 		txq->ring_ref = GRANT_REF_INVALID;
819 		txq->ring.sring = NULL;
820 
821 		snprintf(txq->name, XN_QUEUE_NAME_LEN, "xntx_%u", q);
822 
823 		mtx_init(&txq->lock, txq->name, "netfront transmit lock",
824 		    MTX_DEF);
825 
826 		for (i = 0; i <= NET_TX_RING_SIZE; i++) {
827 			txq->mbufs[i] = (void *) ((u_long) i+1);
828 			txq->grant_ref[i] = GRANT_REF_INVALID;
829 		}
830 		txq->mbufs[NET_TX_RING_SIZE] = (void *)0;
831 
832 		/* Start resources allocation. */
833 
834 		if (gnttab_alloc_grant_references(NET_TX_RING_SIZE,
835 		    &txq->gref_head) != 0) {
836 			device_printf(dev, "failed to allocate tx grant refs\n");
837 			error = ENOMEM;
838 			goto fail;
839 		}
840 
841 		txs = (netif_tx_sring_t *)malloc(PAGE_SIZE, M_DEVBUF,
842 		    M_WAITOK|M_ZERO);
843 		SHARED_RING_INIT(txs);
844 		FRONT_RING_INIT(&txq->ring, txs, PAGE_SIZE);
845 
846 		error = xenbus_grant_ring(dev, virt_to_mfn(txs),
847 		    &txq->ring_ref);
848 		if (error != 0) {
849 			device_printf(dev, "failed to grant tx ring\n");
850 			goto fail_grant_ring;
851 		}
852 
853 		txq->br = buf_ring_alloc(NET_TX_RING_SIZE, M_DEVBUF,
854 		    M_WAITOK, &txq->lock);
855 		TASK_INIT(&txq->defrtask, 0, xn_txq_tq_deferred, txq);
856 
857 		txq->tq = taskqueue_create(txq->name, M_WAITOK,
858 		    taskqueue_thread_enqueue, &txq->tq);
859 
860 		error = taskqueue_start_threads(&txq->tq, 1, PI_NET,
861 		    "%s txq %d", device_get_nameunit(dev), txq->id);
862 		if (error != 0) {
863 			device_printf(dev, "failed to start tx taskq %d\n",
864 			    txq->id);
865 			goto fail_start_thread;
866 		}
867 
868 		error = xen_intr_alloc_and_bind_local_port(dev,
869 		    xenbus_get_otherend_id(dev), /* filter */ NULL, xn_intr,
870 		    &info->txq[q], INTR_TYPE_NET | INTR_MPSAFE | INTR_ENTROPY,
871 		    &txq->xen_intr_handle);
872 
873 		if (error != 0) {
874 			device_printf(dev, "xen_intr_alloc_and_bind_local_port failed\n");
875 			goto fail_bind_port;
876 		}
877 	}
878 
879 	return (0);
880 
881 fail_bind_port:
882 	taskqueue_drain_all(txq->tq);
883 fail_start_thread:
884 	buf_ring_free(txq->br, M_DEVBUF);
885 	taskqueue_free(txq->tq);
886 	gnttab_end_foreign_access(txq->ring_ref, NULL);
887 fail_grant_ring:
888 	gnttab_free_grant_references(txq->gref_head);
889 	free(txq->ring.sring, M_DEVBUF);
890 fail:
891 	for (; q >= 0; q--) {
892 		disconnect_txq(&info->txq[q]);
893 		destroy_txq(&info->txq[q]);
894 	}
895 
896 	free(info->txq, M_DEVBUF);
897 	return (error);
898 }
899 
900 static int
901 setup_device(device_t dev, struct netfront_info *info,
902     unsigned long num_queues)
903 {
904 	int error;
905 	int q;
906 
907 	if (info->txq)
908 		destroy_txqs(info);
909 
910 	if (info->rxq)
911 		destroy_rxqs(info);
912 
913 	info->num_queues = 0;
914 
915 	error = setup_rxqs(dev, info, num_queues);
916 	if (error != 0)
917 		goto out;
918 	error = setup_txqs(dev, info, num_queues);
919 	if (error != 0)
920 		goto out;
921 
922 	info->num_queues = num_queues;
923 
924 	/* No split event channel at the moment. */
925 	for (q = 0; q < num_queues; q++)
926 		info->rxq[q].xen_intr_handle = info->txq[q].xen_intr_handle;
927 
928 	return (0);
929 
930 out:
931 	KASSERT(error != 0, ("Error path taken without providing an error code"));
932 	return (error);
933 }
934 
935 #ifdef INET
936 /**
937  * If this interface has an ipv4 address, send an arp for it. This
938  * helps to get the network going again after migrating hosts.
939  */
940 static void
941 netfront_send_fake_arp(device_t dev, struct netfront_info *info)
942 {
943 	struct ifnet *ifp;
944 	struct ifaddr *ifa;
945 
946 	ifp = info->xn_ifp;
947 	CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) {
948 		if (ifa->ifa_addr->sa_family == AF_INET) {
949 			arp_ifinit(ifp, ifa);
950 		}
951 	}
952 }
953 #endif
954 
955 /**
956  * Callback received when the backend's state changes.
957  */
958 static void
959 netfront_backend_changed(device_t dev, XenbusState newstate)
960 {
961 	struct netfront_info *sc = device_get_softc(dev);
962 
963 	DPRINTK("newstate=%d\n", newstate);
964 
965 	switch (newstate) {
966 	case XenbusStateInitialising:
967 	case XenbusStateInitialised:
968 	case XenbusStateUnknown:
969 	case XenbusStateReconfigured:
970 	case XenbusStateReconfiguring:
971 		break;
972 	case XenbusStateInitWait:
973 		if (xenbus_get_state(dev) != XenbusStateInitialising)
974 			break;
975 		if (xn_connect(sc) != 0)
976 			break;
977 		/* Switch to connected state before kicking the rings. */
978 		xenbus_set_state(sc->xbdev, XenbusStateConnected);
979 		xn_kick_rings(sc);
980 		break;
981 	case XenbusStateClosing:
982 		xenbus_set_state(dev, XenbusStateClosed);
983 		break;
984 	case XenbusStateClosed:
985 		if (sc->xn_reset) {
986 			netif_disconnect_backend(sc);
987 			xenbus_set_state(dev, XenbusStateInitialising);
988 			sc->xn_reset = false;
989 		}
990 		break;
991 	case XenbusStateConnected:
992 #ifdef INET
993 		netfront_send_fake_arp(dev, sc);
994 #endif
995 		break;
996 	}
997 }
998 
999 /**
1000  * \brief Verify that there is sufficient space in the Tx ring
1001  *        buffer for a maximally sized request to be enqueued.
1002  *
1003  * A transmit request requires a transmit descriptor for each packet
1004  * fragment, plus up to 2 entries for "options" (e.g. TSO).
1005  */
1006 static inline int
1007 xn_tx_slot_available(struct netfront_txq *txq)
1008 {
1009 
1010 	return (RING_FREE_REQUESTS(&txq->ring) > (MAX_TX_REQ_FRAGS + 2));
1011 }
1012 
1013 static void
1014 xn_release_tx_bufs(struct netfront_txq *txq)
1015 {
1016 	int i;
1017 
1018 	for (i = 1; i <= NET_TX_RING_SIZE; i++) {
1019 		struct mbuf *m;
1020 
1021 		m = txq->mbufs[i];
1022 
1023 		/*
1024 		 * We assume that no kernel addresses are
1025 		 * less than NET_TX_RING_SIZE.  Any entry
1026 		 * in the table that is below this number
1027 		 * must be an index from free-list tracking.
1028 		 */
1029 		if (((uintptr_t)m) <= NET_TX_RING_SIZE)
1030 			continue;
1031 		gnttab_end_foreign_access_ref(txq->grant_ref[i]);
1032 		gnttab_release_grant_reference(&txq->gref_head,
1033 		    txq->grant_ref[i]);
1034 		txq->grant_ref[i] = GRANT_REF_INVALID;
1035 		add_id_to_freelist(txq->mbufs, i);
1036 		txq->mbufs_cnt--;
1037 		if (txq->mbufs_cnt < 0) {
1038 			panic("%s: tx_chain_cnt must be >= 0", __func__);
1039 		}
1040 		m_free(m);
1041 	}
1042 }
1043 
1044 static struct mbuf *
1045 xn_alloc_one_rx_buffer(struct netfront_rxq *rxq)
1046 {
1047 	struct mbuf *m;
1048 
1049 	m = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MJUMPAGESIZE);
1050 	if (m == NULL)
1051 		return NULL;
1052 	m->m_len = m->m_pkthdr.len = MJUMPAGESIZE;
1053 
1054 	return (m);
1055 }
1056 
1057 static void
1058 xn_alloc_rx_buffers(struct netfront_rxq *rxq)
1059 {
1060 	RING_IDX req_prod;
1061 	int notify;
1062 
1063 	XN_RX_LOCK_ASSERT(rxq);
1064 
1065 	if (__predict_false(rxq->info->carrier == 0))
1066 		return;
1067 
1068 	for (req_prod = rxq->ring.req_prod_pvt;
1069 	     req_prod - rxq->ring.rsp_cons < NET_RX_RING_SIZE;
1070 	     req_prod++) {
1071 		struct mbuf *m;
1072 		unsigned short id;
1073 		grant_ref_t ref;
1074 		struct netif_rx_request *req;
1075 		unsigned long pfn;
1076 
1077 		m = xn_alloc_one_rx_buffer(rxq);
1078 		if (m == NULL)
1079 			break;
1080 
1081 		id = xn_rxidx(req_prod);
1082 
1083 		KASSERT(rxq->mbufs[id] == NULL, ("non-NULL xn_rx_chain"));
1084 		rxq->mbufs[id] = m;
1085 
1086 		ref = gnttab_claim_grant_reference(&rxq->gref_head);
1087 		KASSERT(ref != GNTTAB_LIST_END,
1088 		    ("reserved grant references exhuasted"));
1089 		rxq->grant_ref[id] = ref;
1090 
1091 		pfn = atop(vtophys(mtod(m, vm_offset_t)));
1092 		req = RING_GET_REQUEST(&rxq->ring, req_prod);
1093 
1094 		gnttab_grant_foreign_access_ref(ref,
1095 		    xenbus_get_otherend_id(rxq->info->xbdev), pfn, 0);
1096 		req->id = id;
1097 		req->gref = ref;
1098 	}
1099 
1100 	rxq->ring.req_prod_pvt = req_prod;
1101 
1102 	/* Not enough requests? Try again later. */
1103 	if (req_prod - rxq->ring.rsp_cons < NET_RX_SLOTS_MIN) {
1104 		callout_reset_curcpu(&rxq->rx_refill, hz/10,
1105 		    xn_alloc_rx_buffers_callout, rxq);
1106 		return;
1107 	}
1108 
1109 	wmb();		/* barrier so backend seens requests */
1110 
1111 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&rxq->ring, notify);
1112 	if (notify)
1113 		xen_intr_signal(rxq->xen_intr_handle);
1114 }
1115 
1116 static void xn_alloc_rx_buffers_callout(void *arg)
1117 {
1118 	struct netfront_rxq *rxq;
1119 
1120 	rxq = (struct netfront_rxq *)arg;
1121 	XN_RX_LOCK(rxq);
1122 	xn_alloc_rx_buffers(rxq);
1123 	XN_RX_UNLOCK(rxq);
1124 }
1125 
1126 static void
1127 xn_release_rx_bufs(struct netfront_rxq *rxq)
1128 {
1129 	int i,  ref;
1130 	struct mbuf *m;
1131 
1132 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
1133 		m = rxq->mbufs[i];
1134 
1135 		if (m == NULL)
1136 			continue;
1137 
1138 		ref = rxq->grant_ref[i];
1139 		if (ref == GRANT_REF_INVALID)
1140 			continue;
1141 
1142 		gnttab_end_foreign_access_ref(ref);
1143 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1144 		rxq->mbufs[i] = NULL;
1145 		rxq->grant_ref[i] = GRANT_REF_INVALID;
1146 		m_freem(m);
1147 	}
1148 }
1149 
1150 static void
1151 xn_rxeof(struct netfront_rxq *rxq)
1152 {
1153 	struct ifnet *ifp;
1154 	struct netfront_info *np = rxq->info;
1155 #if (defined(INET) || defined(INET6))
1156 	struct lro_ctrl *lro = &rxq->lro;
1157 #endif
1158 	struct netfront_rx_info rinfo;
1159 	struct netif_rx_response *rx = &rinfo.rx;
1160 	struct netif_extra_info *extras = rinfo.extras;
1161 	RING_IDX i, rp;
1162 	struct mbuf *m;
1163 	struct mbufq mbufq_rxq, mbufq_errq;
1164 	int err, work_to_do;
1165 
1166 	XN_RX_LOCK_ASSERT(rxq);
1167 
1168 	if (!netfront_carrier_ok(np))
1169 		return;
1170 
1171 	/* XXX: there should be some sane limit. */
1172 	mbufq_init(&mbufq_errq, INT_MAX);
1173 	mbufq_init(&mbufq_rxq, INT_MAX);
1174 
1175 	ifp = np->xn_ifp;
1176 
1177 	do {
1178 		rp = rxq->ring.sring->rsp_prod;
1179 		rmb();	/* Ensure we see queued responses up to 'rp'. */
1180 
1181 		i = rxq->ring.rsp_cons;
1182 		while ((i != rp)) {
1183 			memcpy(rx, RING_GET_RESPONSE(&rxq->ring, i), sizeof(*rx));
1184 			memset(extras, 0, sizeof(rinfo.extras));
1185 
1186 			m = NULL;
1187 			err = xn_get_responses(rxq, &rinfo, rp, &i, &m);
1188 
1189 			if (__predict_false(err)) {
1190 				if (m)
1191 					(void )mbufq_enqueue(&mbufq_errq, m);
1192 				if_inc_counter(ifp, IFCOUNTER_IQDROPS, 1);
1193 				continue;
1194 			}
1195 
1196 			m->m_pkthdr.rcvif = ifp;
1197 			if (rx->flags & NETRXF_data_validated) {
1198 				/*
1199 				 * According to mbuf(9) the correct way to tell
1200 				 * the stack that the checksum of an inbound
1201 				 * packet is correct, without it actually being
1202 				 * present (because the underlying interface
1203 				 * doesn't provide it), is to set the
1204 				 * CSUM_DATA_VALID and CSUM_PSEUDO_HDR flags,
1205 				 * and the csum_data field to 0xffff.
1206 				 */
1207 				m->m_pkthdr.csum_flags |= (CSUM_DATA_VALID
1208 				    | CSUM_PSEUDO_HDR);
1209 				m->m_pkthdr.csum_data = 0xffff;
1210 			}
1211 			if ((rx->flags & NETRXF_extra_info) != 0 &&
1212 			    (extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].type ==
1213 			    XEN_NETIF_EXTRA_TYPE_GSO)) {
1214 				m->m_pkthdr.tso_segsz =
1215 				extras[XEN_NETIF_EXTRA_TYPE_GSO - 1].u.gso.size;
1216 				m->m_pkthdr.csum_flags |= CSUM_TSO;
1217 			}
1218 
1219 			(void )mbufq_enqueue(&mbufq_rxq, m);
1220 		}
1221 
1222 		rxq->ring.rsp_cons = i;
1223 
1224 		xn_alloc_rx_buffers(rxq);
1225 
1226 		RING_FINAL_CHECK_FOR_RESPONSES(&rxq->ring, work_to_do);
1227 	} while (work_to_do);
1228 
1229 	mbufq_drain(&mbufq_errq);
1230 	/*
1231 	 * Process all the mbufs after the remapping is complete.
1232 	 * Break the mbuf chain first though.
1233 	 */
1234 	while ((m = mbufq_dequeue(&mbufq_rxq)) != NULL) {
1235 		if_inc_counter(ifp, IFCOUNTER_IPACKETS, 1);
1236 #if (defined(INET) || defined(INET6))
1237 		/* Use LRO if possible */
1238 		if ((ifp->if_capenable & IFCAP_LRO) == 0 ||
1239 		    lro->lro_cnt == 0 || tcp_lro_rx(lro, m, 0)) {
1240 			/*
1241 			 * If LRO fails, pass up to the stack
1242 			 * directly.
1243 			 */
1244 			(*ifp->if_input)(ifp, m);
1245 		}
1246 #else
1247 		(*ifp->if_input)(ifp, m);
1248 #endif
1249 	}
1250 
1251 #if (defined(INET) || defined(INET6))
1252 	/*
1253 	 * Flush any outstanding LRO work
1254 	 */
1255 	tcp_lro_flush_all(lro);
1256 #endif
1257 }
1258 
1259 static void
1260 xn_txeof(struct netfront_txq *txq)
1261 {
1262 	RING_IDX i, prod;
1263 	unsigned short id;
1264 	struct ifnet *ifp;
1265 	netif_tx_response_t *txr;
1266 	struct mbuf *m;
1267 	struct netfront_info *np = txq->info;
1268 
1269 	XN_TX_LOCK_ASSERT(txq);
1270 
1271 	if (!netfront_carrier_ok(np))
1272 		return;
1273 
1274 	ifp = np->xn_ifp;
1275 
1276 	do {
1277 		prod = txq->ring.sring->rsp_prod;
1278 		rmb(); /* Ensure we see responses up to 'rp'. */
1279 
1280 		for (i = txq->ring.rsp_cons; i != prod; i++) {
1281 			txr = RING_GET_RESPONSE(&txq->ring, i);
1282 			if (txr->status == NETIF_RSP_NULL)
1283 				continue;
1284 
1285 			if (txr->status != NETIF_RSP_OKAY) {
1286 				printf("%s: WARNING: response is %d!\n",
1287 				       __func__, txr->status);
1288 			}
1289 			id = txr->id;
1290 			m = txq->mbufs[id];
1291 			KASSERT(m != NULL, ("mbuf not found in chain"));
1292 			KASSERT((uintptr_t)m > NET_TX_RING_SIZE,
1293 				("mbuf already on the free list, but we're "
1294 				"trying to free it again!"));
1295 			M_ASSERTVALID(m);
1296 
1297 			if (__predict_false(gnttab_query_foreign_access(
1298 			    txq->grant_ref[id]) != 0)) {
1299 				panic("%s: grant id %u still in use by the "
1300 				    "backend", __func__, id);
1301 			}
1302 			gnttab_end_foreign_access_ref(txq->grant_ref[id]);
1303 			gnttab_release_grant_reference(
1304 				&txq->gref_head, txq->grant_ref[id]);
1305 			txq->grant_ref[id] = GRANT_REF_INVALID;
1306 
1307 			txq->mbufs[id] = NULL;
1308 			add_id_to_freelist(txq->mbufs, id);
1309 			txq->mbufs_cnt--;
1310 			m_free(m);
1311 			/* Only mark the txq active if we've freed up at least one slot to try */
1312 			ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1313 		}
1314 		txq->ring.rsp_cons = prod;
1315 
1316 		/*
1317 		 * Set a new event, then check for race with update of
1318 		 * tx_cons. Note that it is essential to schedule a
1319 		 * callback, no matter how few buffers are pending. Even if
1320 		 * there is space in the transmit ring, higher layers may
1321 		 * be blocked because too much data is outstanding: in such
1322 		 * cases notification from Xen is likely to be the only kick
1323 		 * that we'll get.
1324 		 */
1325 		txq->ring.sring->rsp_event =
1326 		    prod + ((txq->ring.sring->req_prod - prod) >> 1) + 1;
1327 
1328 		mb();
1329 	} while (prod != txq->ring.sring->rsp_prod);
1330 
1331 	if (txq->full &&
1332 	    ((txq->ring.sring->req_prod - prod) < NET_TX_RING_SIZE)) {
1333 		txq->full = false;
1334 		xn_txq_start(txq);
1335 	}
1336 }
1337 
1338 static void
1339 xn_intr(void *xsc)
1340 {
1341 	struct netfront_txq *txq = xsc;
1342 	struct netfront_info *np = txq->info;
1343 	struct netfront_rxq *rxq = &np->rxq[txq->id];
1344 
1345 	/* kick both tx and rx */
1346 	xn_rxq_intr(rxq);
1347 	xn_txq_intr(txq);
1348 }
1349 
1350 static void
1351 xn_move_rx_slot(struct netfront_rxq *rxq, struct mbuf *m,
1352     grant_ref_t ref)
1353 {
1354 	int new = xn_rxidx(rxq->ring.req_prod_pvt);
1355 
1356 	KASSERT(rxq->mbufs[new] == NULL, ("mbufs != NULL"));
1357 	rxq->mbufs[new] = m;
1358 	rxq->grant_ref[new] = ref;
1359 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->id = new;
1360 	RING_GET_REQUEST(&rxq->ring, rxq->ring.req_prod_pvt)->gref = ref;
1361 	rxq->ring.req_prod_pvt++;
1362 }
1363 
1364 static int
1365 xn_get_extras(struct netfront_rxq *rxq,
1366     struct netif_extra_info *extras, RING_IDX rp, RING_IDX *cons)
1367 {
1368 	struct netif_extra_info *extra;
1369 
1370 	int err = 0;
1371 
1372 	do {
1373 		struct mbuf *m;
1374 		grant_ref_t ref;
1375 
1376 		if (__predict_false(*cons + 1 == rp)) {
1377 			err = EINVAL;
1378 			break;
1379 		}
1380 
1381 		extra = (struct netif_extra_info *)
1382 		RING_GET_RESPONSE(&rxq->ring, ++(*cons));
1383 
1384 		if (__predict_false(!extra->type ||
1385 			extra->type >= XEN_NETIF_EXTRA_TYPE_MAX)) {
1386 			err = EINVAL;
1387 		} else {
1388 			memcpy(&extras[extra->type - 1], extra, sizeof(*extra));
1389 		}
1390 
1391 		m = xn_get_rx_mbuf(rxq, *cons);
1392 		ref = xn_get_rx_ref(rxq,  *cons);
1393 		xn_move_rx_slot(rxq, m, ref);
1394 	} while (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE);
1395 
1396 	return err;
1397 }
1398 
1399 static int
1400 xn_get_responses(struct netfront_rxq *rxq,
1401     struct netfront_rx_info *rinfo, RING_IDX rp, RING_IDX *cons,
1402     struct mbuf  **list)
1403 {
1404 	struct netif_rx_response *rx = &rinfo->rx;
1405 	struct netif_extra_info *extras = rinfo->extras;
1406 	struct mbuf *m, *m0, *m_prev;
1407 	grant_ref_t ref = xn_get_rx_ref(rxq, *cons);
1408 	RING_IDX ref_cons = *cons;
1409 	int frags = 1;
1410 	int err = 0;
1411 	u_long ret;
1412 
1413 	m0 = m = m_prev = xn_get_rx_mbuf(rxq, *cons);
1414 
1415 	if (rx->flags & NETRXF_extra_info) {
1416 		err = xn_get_extras(rxq, extras, rp, cons);
1417 	}
1418 
1419 	if (m0 != NULL) {
1420 		m0->m_pkthdr.len = 0;
1421 		m0->m_next = NULL;
1422 	}
1423 
1424 	for (;;) {
1425 #if 0
1426 		DPRINTK("rx->status=%hd rx->offset=%hu frags=%u\n",
1427 			rx->status, rx->offset, frags);
1428 #endif
1429 		if (__predict_false(rx->status < 0 ||
1430 			rx->offset + rx->status > PAGE_SIZE)) {
1431 
1432 			xn_move_rx_slot(rxq, m, ref);
1433 			if (m0 == m)
1434 				m0 = NULL;
1435 			m = NULL;
1436 			err = EINVAL;
1437 			goto next_skip_queue;
1438 		}
1439 
1440 		/*
1441 		 * This definitely indicates a bug, either in this driver or in
1442 		 * the backend driver. In future this should flag the bad
1443 		 * situation to the system controller to reboot the backed.
1444 		 */
1445 		if (ref == GRANT_REF_INVALID) {
1446 			printf("%s: Bad rx response id %d.\n", __func__, rx->id);
1447 			err = EINVAL;
1448 			goto next;
1449 		}
1450 
1451 		ret = gnttab_end_foreign_access_ref(ref);
1452 		KASSERT(ret, ("Unable to end access to grant references"));
1453 
1454 		gnttab_release_grant_reference(&rxq->gref_head, ref);
1455 
1456 next:
1457 		if (m == NULL)
1458 			break;
1459 
1460 		m->m_len = rx->status;
1461 		m->m_data += rx->offset;
1462 		m0->m_pkthdr.len += rx->status;
1463 
1464 next_skip_queue:
1465 		if (!(rx->flags & NETRXF_more_data))
1466 			break;
1467 
1468 		if (*cons + frags == rp) {
1469 			if (net_ratelimit())
1470 				WPRINTK("Need more frags\n");
1471 			err = ENOENT;
1472 			printf("%s: cons %u frags %u rp %u, not enough frags\n",
1473 			       __func__, *cons, frags, rp);
1474 			break;
1475 		}
1476 		/*
1477 		 * Note that m can be NULL, if rx->status < 0 or if
1478 		 * rx->offset + rx->status > PAGE_SIZE above.
1479 		 */
1480 		m_prev = m;
1481 
1482 		rx = RING_GET_RESPONSE(&rxq->ring, *cons + frags);
1483 		m = xn_get_rx_mbuf(rxq, *cons + frags);
1484 
1485 		/*
1486 		 * m_prev == NULL can happen if rx->status < 0 or if
1487 		 * rx->offset + * rx->status > PAGE_SIZE above.
1488 		 */
1489 		if (m_prev != NULL)
1490 			m_prev->m_next = m;
1491 
1492 		/*
1493 		 * m0 can be NULL if rx->status < 0 or if * rx->offset +
1494 		 * rx->status > PAGE_SIZE above.
1495 		 */
1496 		if (m0 == NULL)
1497 			m0 = m;
1498 		m->m_next = NULL;
1499 		ref = xn_get_rx_ref(rxq, *cons + frags);
1500 		ref_cons = *cons + frags;
1501 		frags++;
1502 	}
1503 	*list = m0;
1504 	*cons += frags;
1505 
1506 	return (err);
1507 }
1508 
1509 /**
1510  * \brief Count the number of fragments in an mbuf chain.
1511  *
1512  * Surprisingly, there isn't an M* macro for this.
1513  */
1514 static inline int
1515 xn_count_frags(struct mbuf *m)
1516 {
1517 	int nfrags;
1518 
1519 	for (nfrags = 0; m != NULL; m = m->m_next)
1520 		nfrags++;
1521 
1522 	return (nfrags);
1523 }
1524 
1525 /**
1526  * Given an mbuf chain, make sure we have enough room and then push
1527  * it onto the transmit ring.
1528  */
1529 static int
1530 xn_assemble_tx_request(struct netfront_txq *txq, struct mbuf *m_head)
1531 {
1532 	struct mbuf *m;
1533 	struct netfront_info *np = txq->info;
1534 	struct ifnet *ifp = np->xn_ifp;
1535 	u_int nfrags;
1536 	int otherend_id;
1537 
1538 	/**
1539 	 * Defragment the mbuf if necessary.
1540 	 */
1541 	nfrags = xn_count_frags(m_head);
1542 
1543 	/*
1544 	 * Check to see whether this request is longer than netback
1545 	 * can handle, and try to defrag it.
1546 	 */
1547 	/**
1548 	 * It is a bit lame, but the netback driver in Linux can't
1549 	 * deal with nfrags > MAX_TX_REQ_FRAGS, which is a quirk of
1550 	 * the Linux network stack.
1551 	 */
1552 	if (nfrags > np->maxfrags) {
1553 		m = m_defrag(m_head, M_NOWAIT);
1554 		if (!m) {
1555 			/*
1556 			 * Defrag failed, so free the mbuf and
1557 			 * therefore drop the packet.
1558 			 */
1559 			m_freem(m_head);
1560 			return (EMSGSIZE);
1561 		}
1562 		m_head = m;
1563 	}
1564 
1565 	/* Determine how many fragments now exist */
1566 	nfrags = xn_count_frags(m_head);
1567 
1568 	/*
1569 	 * Check to see whether the defragmented packet has too many
1570 	 * segments for the Linux netback driver.
1571 	 */
1572 	/**
1573 	 * The FreeBSD TCP stack, with TSO enabled, can produce a chain
1574 	 * of mbufs longer than Linux can handle.  Make sure we don't
1575 	 * pass a too-long chain over to the other side by dropping the
1576 	 * packet.  It doesn't look like there is currently a way to
1577 	 * tell the TCP stack to generate a shorter chain of packets.
1578 	 */
1579 	if (nfrags > MAX_TX_REQ_FRAGS) {
1580 #ifdef DEBUG
1581 		printf("%s: nfrags %d > MAX_TX_REQ_FRAGS %d, netback "
1582 		       "won't be able to handle it, dropping\n",
1583 		       __func__, nfrags, MAX_TX_REQ_FRAGS);
1584 #endif
1585 		m_freem(m_head);
1586 		return (EMSGSIZE);
1587 	}
1588 
1589 	/*
1590 	 * This check should be redundant.  We've already verified that we
1591 	 * have enough slots in the ring to handle a packet of maximum
1592 	 * size, and that our packet is less than the maximum size.  Keep
1593 	 * it in here as an assert for now just to make certain that
1594 	 * chain_cnt is accurate.
1595 	 */
1596 	KASSERT((txq->mbufs_cnt + nfrags) <= NET_TX_RING_SIZE,
1597 		("%s: chain_cnt (%d) + nfrags (%d) > NET_TX_RING_SIZE "
1598 		 "(%d)!", __func__, (int) txq->mbufs_cnt,
1599                     (int) nfrags, (int) NET_TX_RING_SIZE));
1600 
1601 	/*
1602 	 * Start packing the mbufs in this chain into
1603 	 * the fragment pointers. Stop when we run out
1604 	 * of fragments or hit the end of the mbuf chain.
1605 	 */
1606 	m = m_head;
1607 	otherend_id = xenbus_get_otherend_id(np->xbdev);
1608 	for (m = m_head; m; m = m->m_next) {
1609 		netif_tx_request_t *tx;
1610 		uintptr_t id;
1611 		grant_ref_t ref;
1612 		u_long mfn; /* XXX Wrong type? */
1613 
1614 		tx = RING_GET_REQUEST(&txq->ring, txq->ring.req_prod_pvt);
1615 		id = get_id_from_freelist(txq->mbufs);
1616 		if (id == 0)
1617 			panic("%s: was allocated the freelist head!\n",
1618 			    __func__);
1619 		txq->mbufs_cnt++;
1620 		if (txq->mbufs_cnt > NET_TX_RING_SIZE)
1621 			panic("%s: tx_chain_cnt must be <= NET_TX_RING_SIZE\n",
1622 			    __func__);
1623 		txq->mbufs[id] = m;
1624 		tx->id = id;
1625 		ref = gnttab_claim_grant_reference(&txq->gref_head);
1626 		KASSERT((short)ref >= 0, ("Negative ref"));
1627 		mfn = virt_to_mfn(mtod(m, vm_offset_t));
1628 		gnttab_grant_foreign_access_ref(ref, otherend_id,
1629 		    mfn, GNTMAP_readonly);
1630 		tx->gref = txq->grant_ref[id] = ref;
1631 		tx->offset = mtod(m, vm_offset_t) & (PAGE_SIZE - 1);
1632 		tx->flags = 0;
1633 		if (m == m_head) {
1634 			/*
1635 			 * The first fragment has the entire packet
1636 			 * size, subsequent fragments have just the
1637 			 * fragment size. The backend works out the
1638 			 * true size of the first fragment by
1639 			 * subtracting the sizes of the other
1640 			 * fragments.
1641 			 */
1642 			tx->size = m->m_pkthdr.len;
1643 
1644 			/*
1645 			 * The first fragment contains the checksum flags
1646 			 * and is optionally followed by extra data for
1647 			 * TSO etc.
1648 			 */
1649 			/**
1650 			 * CSUM_TSO requires checksum offloading.
1651 			 * Some versions of FreeBSD fail to
1652 			 * set CSUM_TCP in the CSUM_TSO case,
1653 			 * so we have to test for CSUM_TSO
1654 			 * explicitly.
1655 			 */
1656 			if (m->m_pkthdr.csum_flags
1657 			    & (CSUM_DELAY_DATA | CSUM_TSO)) {
1658 				tx->flags |= (NETTXF_csum_blank
1659 				    | NETTXF_data_validated);
1660 			}
1661 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
1662 				struct netif_extra_info *gso =
1663 					(struct netif_extra_info *)
1664 					RING_GET_REQUEST(&txq->ring,
1665 							 ++txq->ring.req_prod_pvt);
1666 
1667 				tx->flags |= NETTXF_extra_info;
1668 
1669 				gso->u.gso.size = m->m_pkthdr.tso_segsz;
1670 				gso->u.gso.type =
1671 					XEN_NETIF_GSO_TYPE_TCPV4;
1672 				gso->u.gso.pad = 0;
1673 				gso->u.gso.features = 0;
1674 
1675 				gso->type = XEN_NETIF_EXTRA_TYPE_GSO;
1676 				gso->flags = 0;
1677 			}
1678 		} else {
1679 			tx->size = m->m_len;
1680 		}
1681 		if (m->m_next)
1682 			tx->flags |= NETTXF_more_data;
1683 
1684 		txq->ring.req_prod_pvt++;
1685 	}
1686 	BPF_MTAP(ifp, m_head);
1687 
1688 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, 1);
1689 	if_inc_counter(ifp, IFCOUNTER_OBYTES, m_head->m_pkthdr.len);
1690 	if (m_head->m_flags & M_MCAST)
1691 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, 1);
1692 
1693 	xn_txeof(txq);
1694 
1695 	return (0);
1696 }
1697 
1698 /* equivalent of network_open() in Linux */
1699 static void
1700 xn_ifinit_locked(struct netfront_info *np)
1701 {
1702 	struct ifnet *ifp;
1703 	int i;
1704 	struct netfront_rxq *rxq;
1705 
1706 	XN_LOCK_ASSERT(np);
1707 
1708 	ifp = np->xn_ifp;
1709 
1710 	if (ifp->if_drv_flags & IFF_DRV_RUNNING || !netfront_carrier_ok(np))
1711 		return;
1712 
1713 	xn_stop(np);
1714 
1715 	for (i = 0; i < np->num_queues; i++) {
1716 		rxq = &np->rxq[i];
1717 		XN_RX_LOCK(rxq);
1718 		xn_alloc_rx_buffers(rxq);
1719 		rxq->ring.sring->rsp_event = rxq->ring.rsp_cons + 1;
1720 		if (RING_HAS_UNCONSUMED_RESPONSES(&rxq->ring))
1721 			xn_rxeof(rxq);
1722 		XN_RX_UNLOCK(rxq);
1723 	}
1724 
1725 	ifp->if_drv_flags |= IFF_DRV_RUNNING;
1726 	ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
1727 	if_link_state_change(ifp, LINK_STATE_UP);
1728 }
1729 
1730 static void
1731 xn_ifinit(void *xsc)
1732 {
1733 	struct netfront_info *sc = xsc;
1734 
1735 	XN_LOCK(sc);
1736 	xn_ifinit_locked(sc);
1737 	XN_UNLOCK(sc);
1738 }
1739 
1740 static int
1741 xn_ioctl(struct ifnet *ifp, u_long cmd, caddr_t data)
1742 {
1743 	struct netfront_info *sc = ifp->if_softc;
1744 	struct ifreq *ifr = (struct ifreq *) data;
1745 	device_t dev;
1746 #ifdef INET
1747 	struct ifaddr *ifa = (struct ifaddr *)data;
1748 #endif
1749 	int mask, error = 0, reinit;
1750 
1751 	dev = sc->xbdev;
1752 
1753 	switch(cmd) {
1754 	case SIOCSIFADDR:
1755 #ifdef INET
1756 		XN_LOCK(sc);
1757 		if (ifa->ifa_addr->sa_family == AF_INET) {
1758 			ifp->if_flags |= IFF_UP;
1759 			if (!(ifp->if_drv_flags & IFF_DRV_RUNNING))
1760 				xn_ifinit_locked(sc);
1761 			arp_ifinit(ifp, ifa);
1762 			XN_UNLOCK(sc);
1763 		} else {
1764 			XN_UNLOCK(sc);
1765 #endif
1766 			error = ether_ioctl(ifp, cmd, data);
1767 #ifdef INET
1768 		}
1769 #endif
1770 		break;
1771 	case SIOCSIFMTU:
1772 		if (ifp->if_mtu == ifr->ifr_mtu)
1773 			break;
1774 
1775 		ifp->if_mtu = ifr->ifr_mtu;
1776 		ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
1777 		xn_ifinit(sc);
1778 		break;
1779 	case SIOCSIFFLAGS:
1780 		XN_LOCK(sc);
1781 		if (ifp->if_flags & IFF_UP) {
1782 			/*
1783 			 * If only the state of the PROMISC flag changed,
1784 			 * then just use the 'set promisc mode' command
1785 			 * instead of reinitializing the entire NIC. Doing
1786 			 * a full re-init means reloading the firmware and
1787 			 * waiting for it to start up, which may take a
1788 			 * second or two.
1789 			 */
1790 			xn_ifinit_locked(sc);
1791 		} else {
1792 			if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
1793 				xn_stop(sc);
1794 			}
1795 		}
1796 		sc->xn_if_flags = ifp->if_flags;
1797 		XN_UNLOCK(sc);
1798 		break;
1799 	case SIOCSIFCAP:
1800 		mask = ifr->ifr_reqcap ^ ifp->if_capenable;
1801 		reinit = 0;
1802 
1803 		if (mask & IFCAP_TXCSUM) {
1804 			ifp->if_capenable ^= IFCAP_TXCSUM;
1805 			ifp->if_hwassist ^= XN_CSUM_FEATURES;
1806 		}
1807 		if (mask & IFCAP_TSO4) {
1808 			ifp->if_capenable ^= IFCAP_TSO4;
1809 			ifp->if_hwassist ^= CSUM_TSO;
1810 		}
1811 
1812 		if (mask & (IFCAP_RXCSUM | IFCAP_LRO)) {
1813 			/* These Rx features require us to renegotiate. */
1814 			reinit = 1;
1815 
1816 			if (mask & IFCAP_RXCSUM)
1817 				ifp->if_capenable ^= IFCAP_RXCSUM;
1818 			if (mask & IFCAP_LRO)
1819 				ifp->if_capenable ^= IFCAP_LRO;
1820 		}
1821 
1822 		if (reinit == 0)
1823 			break;
1824 
1825 		/*
1826 		 * We must reset the interface so the backend picks up the
1827 		 * new features.
1828 		 */
1829 		device_printf(sc->xbdev,
1830 		    "performing interface reset due to feature change\n");
1831 		XN_LOCK(sc);
1832 		netfront_carrier_off(sc);
1833 		sc->xn_reset = true;
1834 		/*
1835 		 * NB: the pending packet queue is not flushed, since
1836 		 * the interface should still support the old options.
1837 		 */
1838 		XN_UNLOCK(sc);
1839 		/*
1840 		 * Delete the xenstore nodes that export features.
1841 		 *
1842 		 * NB: There's a xenbus state called
1843 		 * "XenbusStateReconfiguring", which is what we should set
1844 		 * here. Sadly none of the backends know how to handle it,
1845 		 * and simply disconnect from the frontend, so we will just
1846 		 * switch back to XenbusStateInitialising in order to force
1847 		 * a reconnection.
1848 		 */
1849 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-gso-tcpv4");
1850 		xs_rm(XST_NIL, xenbus_get_node(dev), "feature-no-csum-offload");
1851 		xenbus_set_state(dev, XenbusStateClosing);
1852 
1853 		/*
1854 		 * Wait for the frontend to reconnect before returning
1855 		 * from the ioctl. 30s should be more than enough for any
1856 		 * sane backend to reconnect.
1857 		 */
1858 		error = tsleep(sc, 0, "xn_rst", 30*hz);
1859 		break;
1860 	case SIOCADDMULTI:
1861 	case SIOCDELMULTI:
1862 		break;
1863 	case SIOCSIFMEDIA:
1864 	case SIOCGIFMEDIA:
1865 		error = ifmedia_ioctl(ifp, ifr, &sc->sc_media, cmd);
1866 		break;
1867 	default:
1868 		error = ether_ioctl(ifp, cmd, data);
1869 	}
1870 
1871 	return (error);
1872 }
1873 
1874 static void
1875 xn_stop(struct netfront_info *sc)
1876 {
1877 	struct ifnet *ifp;
1878 
1879 	XN_LOCK_ASSERT(sc);
1880 
1881 	ifp = sc->xn_ifp;
1882 
1883 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
1884 	if_link_state_change(ifp, LINK_STATE_DOWN);
1885 }
1886 
1887 static void
1888 xn_rebuild_rx_bufs(struct netfront_rxq *rxq)
1889 {
1890 	int requeue_idx, i;
1891 	grant_ref_t ref;
1892 	netif_rx_request_t *req;
1893 
1894 	for (requeue_idx = 0, i = 0; i < NET_RX_RING_SIZE; i++) {
1895 		struct mbuf *m;
1896 		u_long pfn;
1897 
1898 		if (rxq->mbufs[i] == NULL)
1899 			continue;
1900 
1901 		m = rxq->mbufs[requeue_idx] = xn_get_rx_mbuf(rxq, i);
1902 		ref = rxq->grant_ref[requeue_idx] = xn_get_rx_ref(rxq, i);
1903 
1904 		req = RING_GET_REQUEST(&rxq->ring, requeue_idx);
1905 		pfn = vtophys(mtod(m, vm_offset_t)) >> PAGE_SHIFT;
1906 
1907 		gnttab_grant_foreign_access_ref(ref,
1908 		    xenbus_get_otherend_id(rxq->info->xbdev),
1909 		    pfn, 0);
1910 
1911 		req->gref = ref;
1912 		req->id   = requeue_idx;
1913 
1914 		requeue_idx++;
1915 	}
1916 
1917 	rxq->ring.req_prod_pvt = requeue_idx;
1918 }
1919 
1920 /* START of Xenolinux helper functions adapted to FreeBSD */
1921 static int
1922 xn_connect(struct netfront_info *np)
1923 {
1924 	int i, error;
1925 	u_int feature_rx_copy;
1926 	struct netfront_rxq *rxq;
1927 	struct netfront_txq *txq;
1928 
1929 	error = xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1930 	    "feature-rx-copy", NULL, "%u", &feature_rx_copy);
1931 	if (error != 0)
1932 		feature_rx_copy = 0;
1933 
1934 	/* We only support rx copy. */
1935 	if (!feature_rx_copy)
1936 		return (EPROTONOSUPPORT);
1937 
1938 	/* Recovery procedure: */
1939 	error = talk_to_backend(np->xbdev, np);
1940 	if (error != 0)
1941 		return (error);
1942 
1943 	/* Step 1: Reinitialise variables. */
1944 	xn_query_features(np);
1945 	xn_configure_features(np);
1946 
1947 	/* Step 2: Release TX buffer */
1948 	for (i = 0; i < np->num_queues; i++) {
1949 		txq = &np->txq[i];
1950 		xn_release_tx_bufs(txq);
1951 	}
1952 
1953 	/* Step 3: Rebuild the RX buffer freelist and the RX ring itself. */
1954 	for (i = 0; i < np->num_queues; i++) {
1955 		rxq = &np->rxq[i];
1956 		xn_rebuild_rx_bufs(rxq);
1957 	}
1958 
1959 	/* Step 4: All public and private state should now be sane.  Get
1960 	 * ready to start sending and receiving packets and give the driver
1961 	 * domain a kick because we've probably just requeued some
1962 	 * packets.
1963 	 */
1964 	netfront_carrier_on(np);
1965 	wakeup(np);
1966 
1967 	return (0);
1968 }
1969 
1970 static void
1971 xn_kick_rings(struct netfront_info *np)
1972 {
1973 	struct netfront_rxq *rxq;
1974 	struct netfront_txq *txq;
1975 	int i;
1976 
1977 	for (i = 0; i < np->num_queues; i++) {
1978 		txq = &np->txq[i];
1979 		rxq = &np->rxq[i];
1980 		xen_intr_signal(txq->xen_intr_handle);
1981 		XN_TX_LOCK(txq);
1982 		xn_txeof(txq);
1983 		XN_TX_UNLOCK(txq);
1984 		XN_RX_LOCK(rxq);
1985 		xn_alloc_rx_buffers(rxq);
1986 		XN_RX_UNLOCK(rxq);
1987 	}
1988 }
1989 
1990 static void
1991 xn_query_features(struct netfront_info *np)
1992 {
1993 	int val;
1994 
1995 	device_printf(np->xbdev, "backend features:");
1996 
1997 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
1998 		"feature-sg", NULL, "%d", &val) != 0)
1999 		val = 0;
2000 
2001 	np->maxfrags = 1;
2002 	if (val) {
2003 		np->maxfrags = MAX_TX_REQ_FRAGS;
2004 		printf(" feature-sg");
2005 	}
2006 
2007 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2008 		"feature-gso-tcpv4", NULL, "%d", &val) != 0)
2009 		val = 0;
2010 
2011 	np->xn_ifp->if_capabilities &= ~(IFCAP_TSO4|IFCAP_LRO);
2012 	if (val) {
2013 		np->xn_ifp->if_capabilities |= IFCAP_TSO4|IFCAP_LRO;
2014 		printf(" feature-gso-tcp4");
2015 	}
2016 
2017 	/*
2018 	 * HW CSUM offload is assumed to be available unless
2019 	 * feature-no-csum-offload is set in xenstore.
2020 	 */
2021 	if (xs_scanf(XST_NIL, xenbus_get_otherend_path(np->xbdev),
2022 		"feature-no-csum-offload", NULL, "%d", &val) != 0)
2023 		val = 0;
2024 
2025 	np->xn_ifp->if_capabilities |= IFCAP_HWCSUM;
2026 	if (val) {
2027 		np->xn_ifp->if_capabilities &= ~(IFCAP_HWCSUM);
2028 		printf(" feature-no-csum-offload");
2029 	}
2030 
2031 	printf("\n");
2032 }
2033 
2034 static int
2035 xn_configure_features(struct netfront_info *np)
2036 {
2037 	int err, cap_enabled;
2038 #if (defined(INET) || defined(INET6))
2039 	int i;
2040 #endif
2041 	struct ifnet *ifp;
2042 
2043 	ifp = np->xn_ifp;
2044 	err = 0;
2045 
2046 	if ((ifp->if_capenable & ifp->if_capabilities) == ifp->if_capenable) {
2047 		/* Current options are available, no need to do anything. */
2048 		return (0);
2049 	}
2050 
2051 	/* Try to preserve as many options as possible. */
2052 	cap_enabled = ifp->if_capenable;
2053 	ifp->if_capenable = ifp->if_hwassist = 0;
2054 
2055 #if (defined(INET) || defined(INET6))
2056 	if ((cap_enabled & IFCAP_LRO) != 0)
2057 		for (i = 0; i < np->num_queues; i++)
2058 			tcp_lro_free(&np->rxq[i].lro);
2059 	if (xn_enable_lro &&
2060 	    (ifp->if_capabilities & cap_enabled & IFCAP_LRO) != 0) {
2061 	    	ifp->if_capenable |= IFCAP_LRO;
2062 		for (i = 0; i < np->num_queues; i++) {
2063 			err = tcp_lro_init(&np->rxq[i].lro);
2064 			if (err != 0) {
2065 				device_printf(np->xbdev,
2066 				    "LRO initialization failed\n");
2067 				ifp->if_capenable &= ~IFCAP_LRO;
2068 				break;
2069 			}
2070 			np->rxq[i].lro.ifp = ifp;
2071 		}
2072 	}
2073 	if ((ifp->if_capabilities & cap_enabled & IFCAP_TSO4) != 0) {
2074 		ifp->if_capenable |= IFCAP_TSO4;
2075 		ifp->if_hwassist |= CSUM_TSO;
2076 	}
2077 #endif
2078 	if ((ifp->if_capabilities & cap_enabled & IFCAP_TXCSUM) != 0) {
2079 		ifp->if_capenable |= IFCAP_TXCSUM;
2080 		ifp->if_hwassist |= XN_CSUM_FEATURES;
2081 	}
2082 	if ((ifp->if_capabilities & cap_enabled & IFCAP_RXCSUM) != 0)
2083 		ifp->if_capenable |= IFCAP_RXCSUM;
2084 
2085 	return (err);
2086 }
2087 
2088 static int
2089 xn_txq_mq_start_locked(struct netfront_txq *txq, struct mbuf *m)
2090 {
2091 	struct netfront_info *np;
2092 	struct ifnet *ifp;
2093 	struct buf_ring *br;
2094 	int error, notify;
2095 
2096 	np = txq->info;
2097 	br = txq->br;
2098 	ifp = np->xn_ifp;
2099 	error = 0;
2100 
2101 	XN_TX_LOCK_ASSERT(txq);
2102 
2103 	if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 ||
2104 	    !netfront_carrier_ok(np)) {
2105 		if (m != NULL)
2106 			error = drbr_enqueue(ifp, br, m);
2107 		return (error);
2108 	}
2109 
2110 	if (m != NULL) {
2111 		error = drbr_enqueue(ifp, br, m);
2112 		if (error != 0)
2113 			return (error);
2114 	}
2115 
2116 	while ((m = drbr_peek(ifp, br)) != NULL) {
2117 		if (!xn_tx_slot_available(txq)) {
2118 			drbr_putback(ifp, br, m);
2119 			break;
2120 		}
2121 
2122 		error = xn_assemble_tx_request(txq, m);
2123 		/* xn_assemble_tx_request always consumes the mbuf*/
2124 		if (error != 0) {
2125 			drbr_advance(ifp, br);
2126 			break;
2127 		}
2128 
2129 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&txq->ring, notify);
2130 		if (notify)
2131 			xen_intr_signal(txq->xen_intr_handle);
2132 
2133 		drbr_advance(ifp, br);
2134 	}
2135 
2136 	if (RING_FULL(&txq->ring))
2137 		txq->full = true;
2138 
2139 	return (0);
2140 }
2141 
2142 static int
2143 xn_txq_mq_start(struct ifnet *ifp, struct mbuf *m)
2144 {
2145 	struct netfront_info *np;
2146 	struct netfront_txq *txq;
2147 	int i, npairs, error;
2148 
2149 	np = ifp->if_softc;
2150 	npairs = np->num_queues;
2151 
2152 	if (!netfront_carrier_ok(np))
2153 		return (ENOBUFS);
2154 
2155 	KASSERT(npairs != 0, ("called with 0 available queues"));
2156 
2157 	/* check if flowid is set */
2158 	if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE)
2159 		i = m->m_pkthdr.flowid % npairs;
2160 	else
2161 		i = curcpu % npairs;
2162 
2163 	txq = &np->txq[i];
2164 
2165 	if (XN_TX_TRYLOCK(txq) != 0) {
2166 		error = xn_txq_mq_start_locked(txq, m);
2167 		XN_TX_UNLOCK(txq);
2168 	} else {
2169 		error = drbr_enqueue(ifp, txq->br, m);
2170 		taskqueue_enqueue(txq->tq, &txq->defrtask);
2171 	}
2172 
2173 	return (error);
2174 }
2175 
2176 static void
2177 xn_qflush(struct ifnet *ifp)
2178 {
2179 	struct netfront_info *np;
2180 	struct netfront_txq *txq;
2181 	struct mbuf *m;
2182 	int i;
2183 
2184 	np = ifp->if_softc;
2185 
2186 	for (i = 0; i < np->num_queues; i++) {
2187 		txq = &np->txq[i];
2188 
2189 		XN_TX_LOCK(txq);
2190 		while ((m = buf_ring_dequeue_sc(txq->br)) != NULL)
2191 			m_freem(m);
2192 		XN_TX_UNLOCK(txq);
2193 	}
2194 
2195 	if_qflush(ifp);
2196 }
2197 
2198 /**
2199  * Create a network device.
2200  * @param dev  Newbus device representing this virtual NIC.
2201  */
2202 int
2203 create_netdev(device_t dev)
2204 {
2205 	struct netfront_info *np;
2206 	int err;
2207 	struct ifnet *ifp;
2208 
2209 	np = device_get_softc(dev);
2210 
2211 	np->xbdev         = dev;
2212 
2213 	mtx_init(&np->sc_lock, "xnsc", "netfront softc lock", MTX_DEF);
2214 
2215 	ifmedia_init(&np->sc_media, 0, xn_ifmedia_upd, xn_ifmedia_sts);
2216 	ifmedia_add(&np->sc_media, IFM_ETHER|IFM_MANUAL, 0, NULL);
2217 	ifmedia_set(&np->sc_media, IFM_ETHER|IFM_MANUAL);
2218 
2219 	err = xen_net_read_mac(dev, np->mac);
2220 	if (err != 0)
2221 		goto error;
2222 
2223 	/* Set up ifnet structure */
2224 	ifp = np->xn_ifp = if_alloc(IFT_ETHER);
2225     	ifp->if_softc = np;
2226     	if_initname(ifp, "xn",  device_get_unit(dev));
2227     	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
2228     	ifp->if_ioctl = xn_ioctl;
2229 
2230 	ifp->if_transmit = xn_txq_mq_start;
2231 	ifp->if_qflush = xn_qflush;
2232 
2233     	ifp->if_init = xn_ifinit;
2234 
2235     	ifp->if_hwassist = XN_CSUM_FEATURES;
2236 	/* Enable all supported features at device creation. */
2237 	ifp->if_capenable = ifp->if_capabilities =
2238 	    IFCAP_HWCSUM|IFCAP_TSO4|IFCAP_LRO;
2239 	ifp->if_hw_tsomax = 65536 - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
2240 	ifp->if_hw_tsomaxsegcount = MAX_TX_REQ_FRAGS;
2241 	ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
2242 
2243     	ether_ifattach(ifp, np->mac);
2244 	netfront_carrier_off(np);
2245 
2246 	return (0);
2247 
2248 error:
2249 	KASSERT(err != 0, ("Error path with no error code specified"));
2250 	return (err);
2251 }
2252 
2253 static int
2254 netfront_detach(device_t dev)
2255 {
2256 	struct netfront_info *info = device_get_softc(dev);
2257 
2258 	DPRINTK("%s\n", xenbus_get_node(dev));
2259 
2260 	netif_free(info);
2261 
2262 	return 0;
2263 }
2264 
2265 static void
2266 netif_free(struct netfront_info *np)
2267 {
2268 
2269 	XN_LOCK(np);
2270 	xn_stop(np);
2271 	XN_UNLOCK(np);
2272 	netif_disconnect_backend(np);
2273 	ether_ifdetach(np->xn_ifp);
2274 	free(np->rxq, M_DEVBUF);
2275 	free(np->txq, M_DEVBUF);
2276 	if_free(np->xn_ifp);
2277 	np->xn_ifp = NULL;
2278 	ifmedia_removeall(&np->sc_media);
2279 }
2280 
2281 static void
2282 netif_disconnect_backend(struct netfront_info *np)
2283 {
2284 	u_int i;
2285 
2286 	for (i = 0; i < np->num_queues; i++) {
2287 		XN_RX_LOCK(&np->rxq[i]);
2288 		XN_TX_LOCK(&np->txq[i]);
2289 	}
2290 	netfront_carrier_off(np);
2291 	for (i = 0; i < np->num_queues; i++) {
2292 		XN_RX_UNLOCK(&np->rxq[i]);
2293 		XN_TX_UNLOCK(&np->txq[i]);
2294 	}
2295 
2296 	for (i = 0; i < np->num_queues; i++) {
2297 		disconnect_rxq(&np->rxq[i]);
2298 		disconnect_txq(&np->txq[i]);
2299 	}
2300 }
2301 
2302 static int
2303 xn_ifmedia_upd(struct ifnet *ifp)
2304 {
2305 
2306 	return (0);
2307 }
2308 
2309 static void
2310 xn_ifmedia_sts(struct ifnet *ifp, struct ifmediareq *ifmr)
2311 {
2312 
2313 	ifmr->ifm_status = IFM_AVALID|IFM_ACTIVE;
2314 	ifmr->ifm_active = IFM_ETHER|IFM_MANUAL;
2315 }
2316 
2317 /* ** Driver registration ** */
2318 static device_method_t netfront_methods[] = {
2319 	/* Device interface */
2320 	DEVMETHOD(device_probe,         netfront_probe),
2321 	DEVMETHOD(device_attach,        netfront_attach),
2322 	DEVMETHOD(device_detach,        netfront_detach),
2323 	DEVMETHOD(device_shutdown,      bus_generic_shutdown),
2324 	DEVMETHOD(device_suspend,       netfront_suspend),
2325 	DEVMETHOD(device_resume,        netfront_resume),
2326 
2327 	/* Xenbus interface */
2328 	DEVMETHOD(xenbus_otherend_changed, netfront_backend_changed),
2329 
2330 	DEVMETHOD_END
2331 };
2332 
2333 static driver_t netfront_driver = {
2334 	"xn",
2335 	netfront_methods,
2336 	sizeof(struct netfront_info),
2337 };
2338 devclass_t netfront_devclass;
2339 
2340 DRIVER_MODULE(xe, xenbusb_front, netfront_driver, netfront_devclass, NULL,
2341     NULL);
2342