xref: /illumos-gate/usr/src/uts/common/xen/io/xnf.c (revision 0b4d657578b8f707371421ebae5b3837023b4144)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Copyright (c) 2014, 2017 by Delphix. All rights reserved.
29  */
30 
31 /*
32  *
33  * Copyright (c) 2004 Christian Limpach.
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  * 3. This section intentionally left blank.
45  * 4. The name of the author may not be used to endorse or promote products
46  *    derived from this software without specific prior written permission.
47  *
48  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
49  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
50  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
51  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
52  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
53  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
54  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
55  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
56  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
57  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
58  */
59 /*
60  * Section 3 of the above license was updated in response to bug 6379571.
61  */
62 
63 /*
64  * xnf.c - GLDv3 network driver for domU.
65  */
66 
67 /*
68  * This driver uses four per-instance locks:
69  *
70  * xnf_gref_lock:
71  *
72  *    Protects access to the grant reference list stored in
73  *    xnf_gref_head. Grant references should be acquired and released
74  *    using gref_get() and gref_put() respectively.
75  *
76  * xnf_schedlock:
77  *
78  *    Protects:
79  *    xnf_need_sched - used to record that a previous transmit attempt
80  *       failed (and consequently it will be necessary to call
81  *       mac_tx_update() when transmit resources are available).
82  *    xnf_pending_multicast - the number of multicast requests that
83  *       have been submitted to the backend for which we have not
84  *       processed responses.
85  *
86  * xnf_txlock:
87  *
88  *    Protects the transmit ring (xnf_tx_ring) and associated
89  *    structures (notably xnf_tx_pkt_id and xnf_tx_pkt_id_head).
90  *
91  * xnf_rxlock:
92  *
93  *    Protects the receive ring (xnf_rx_ring) and associated
94  *    structures (notably xnf_rx_pkt_info).
95  *
96  * If driver-global state that affects both the transmit and receive
97  * rings is manipulated, both xnf_txlock and xnf_rxlock should be
98  * held, in that order.
99  *
100  * xnf_schedlock is acquired both whilst holding xnf_txlock and
101  * without. It should always be acquired after xnf_txlock if both are
102  * held.
103  *
104  * Notes:
105  * - atomic_add_64() is used to manipulate counters where we require
106  *   accuracy. For counters intended only for observation by humans,
107  *   post increment/decrement are used instead.
108  */
109 
110 #include <sys/types.h>
111 #include <sys/errno.h>
112 #include <sys/param.h>
113 #include <sys/sysmacros.h>
114 #include <sys/systm.h>
115 #include <sys/stream.h>
116 #include <sys/strsubr.h>
117 #include <sys/strsun.h>
118 #include <sys/conf.h>
119 #include <sys/ddi.h>
120 #include <sys/devops.h>
121 #include <sys/sunddi.h>
122 #include <sys/sunndi.h>
123 #include <sys/dlpi.h>
124 #include <sys/ethernet.h>
125 #include <sys/strsun.h>
126 #include <sys/pattr.h>
127 #include <inet/ip.h>
128 #include <inet/ip_impl.h>
129 #include <inet/tcp.h>
130 #include <netinet/udp.h>
131 #include <sys/gld.h>
132 #include <sys/modctl.h>
133 #include <sys/mac_provider.h>
134 #include <sys/mac_ether.h>
135 #include <sys/bootinfo.h>
136 #include <sys/mach_mmu.h>
137 #ifdef	XPV_HVM_DRIVER
138 #include <sys/xpv_support.h>
139 #include <sys/hypervisor.h>
140 #else
141 #include <sys/hypervisor.h>
142 #include <sys/evtchn_impl.h>
143 #include <sys/balloon_impl.h>
144 #endif
145 #include <xen/public/io/netif.h>
146 #include <sys/gnttab.h>
147 #include <xen/sys/xendev.h>
148 #include <sys/sdt.h>
149 #include <sys/note.h>
150 #include <sys/debug.h>
151 
152 #include <io/xnf.h>
153 
154 #if defined(DEBUG) || defined(__lint)
155 #define	XNF_DEBUG
156 #endif
157 
158 #ifdef XNF_DEBUG
159 int xnf_debug = 0;
160 xnf_t *xnf_debug_instance = NULL;
161 #endif
162 
163 /*
164  * On a 32 bit PAE system physical and machine addresses are larger
165  * than 32 bits.  ddi_btop() on such systems take an unsigned long
166  * argument, and so addresses above 4G are truncated before ddi_btop()
167  * gets to see them.  To avoid this, code the shift operation here.
168  */
169 #define	xnf_btop(addr)	((addr) >> PAGESHIFT)
170 
171 /*
172  * The parameters below should only be changed in /etc/system, never in mdb.
173  */
174 
175 /*
176  * Should we use the multicast control feature if the backend provides
177  * it?
178  */
179 boolean_t xnf_multicast_control = B_TRUE;
180 
181 /*
182  * Should we allow scatter-gather for tx if backend allows it?
183  */
184 boolean_t xnf_enable_tx_sg = B_TRUE;
185 
186 /*
187  * Should we allow scatter-gather for rx if backend allows it?
188  */
189 boolean_t xnf_enable_rx_sg = B_TRUE;
190 
191 /*
192  * Should we allow lso for tx sends if backend allows it?
193  * Requires xnf_enable_tx_sg to be also set to TRUE.
194  */
195 boolean_t xnf_enable_lso = B_TRUE;
196 
197 /*
198  * Should we allow lro on rx if backend supports it?
199  * Requires xnf_enable_rx_sg to be also set to TRUE.
200  *
201  * !! WARNING !!
202  * LRO is not yet supported in the OS so this should be left as FALSE.
203  * !! WARNING !!
204  */
205 boolean_t xnf_enable_lro = B_FALSE;
206 
207 /*
208  * Received packets below this size are copied to a new streams buffer
209  * rather than being desballoc'ed.
210  *
211  * This value is chosen to accommodate traffic where there are a large
212  * number of small packets. For data showing a typical distribution,
213  * see:
214  *
215  * Sinha07a:
216  *	Rishi Sinha, Christos Papadopoulos, and John
217  *	Heidemann. Internet Packet Size Distributions: Some
218  *	Observations. Technical Report ISI-TR-2007-643,
219  *	USC/Information Sciences Institute, May, 2007. Orignally
220  *	released October 2005 as web page
221  *	http://netweb.usc.edu/~sinha/pkt-sizes/.
222  *	<http://www.isi.edu/~johnh/PAPERS/Sinha07a.html>.
223  */
224 size_t xnf_rx_copy_limit = 64;
225 
226 #define	INVALID_GRANT_HANDLE	((grant_handle_t)-1)
227 #define	INVALID_GRANT_REF	((grant_ref_t)-1)
228 #define	INVALID_TX_ID		((uint16_t)-1)
229 
230 #define	TX_ID_TO_TXID(p, id) (&((p)->xnf_tx_pkt_id[(id)]))
231 #define	TX_ID_VALID(i) \
232 	(((i) != INVALID_TX_ID) && ((i) < NET_TX_RING_SIZE))
233 
234 /*
235  * calculate how many pages are spanned by an mblk fragment
236  */
237 #define	xnf_mblk_pages(mp)	(MBLKL(mp) == 0 ? 0 : \
238     xnf_btop((uintptr_t)mp->b_wptr - 1) - xnf_btop((uintptr_t)mp->b_rptr) + 1)
239 
240 /* Required system entry points */
241 static int	xnf_attach(dev_info_t *, ddi_attach_cmd_t);
242 static int	xnf_detach(dev_info_t *, ddi_detach_cmd_t);
243 
244 /* Required driver entry points for Nemo */
245 static int	xnf_start(void *);
246 static void	xnf_stop(void *);
247 static int	xnf_set_mac_addr(void *, const uint8_t *);
248 static int	xnf_set_multicast(void *, boolean_t, const uint8_t *);
249 static int	xnf_set_promiscuous(void *, boolean_t);
250 static mblk_t	*xnf_send(void *, mblk_t *);
251 static uint_t	xnf_intr(caddr_t);
252 static int	xnf_stat(void *, uint_t, uint64_t *);
253 static boolean_t xnf_getcapab(void *, mac_capab_t, void *);
254 static int xnf_getprop(void *, const char *, mac_prop_id_t, uint_t, void *);
255 static int xnf_setprop(void *, const char *, mac_prop_id_t, uint_t,
256     const void *);
257 static void xnf_propinfo(void *, const char *, mac_prop_id_t,
258     mac_prop_info_handle_t);
259 
260 /* Driver private functions */
261 static int xnf_alloc_dma_resources(xnf_t *);
262 static void xnf_release_dma_resources(xnf_t *);
263 static void xnf_release_mblks(xnf_t *);
264 
265 static int xnf_buf_constructor(void *, void *, int);
266 static void xnf_buf_destructor(void *, void *);
267 static xnf_buf_t *xnf_buf_get(xnf_t *, int, boolean_t);
268 #pragma inline(xnf_buf_get)
269 static void xnf_buf_put(xnf_t *, xnf_buf_t *, boolean_t);
270 #pragma inline(xnf_buf_put)
271 static void xnf_buf_refresh(xnf_buf_t *);
272 #pragma inline(xnf_buf_refresh)
273 static void xnf_buf_recycle(xnf_buf_t *);
274 
275 static int xnf_tx_buf_constructor(void *, void *, int);
276 static void xnf_tx_buf_destructor(void *, void *);
277 
278 static grant_ref_t xnf_gref_get(xnf_t *);
279 #pragma inline(xnf_gref_get)
280 static void xnf_gref_put(xnf_t *, grant_ref_t);
281 #pragma inline(xnf_gref_put)
282 
283 static xnf_txid_t *xnf_txid_get(xnf_t *);
284 #pragma inline(xnf_txid_get)
285 static void xnf_txid_put(xnf_t *, xnf_txid_t *);
286 #pragma inline(xnf_txid_put)
287 
288 static void xnf_rxbuf_hang(xnf_t *, xnf_buf_t *);
289 static int xnf_tx_clean_ring(xnf_t  *);
290 static void oe_state_change(dev_info_t *, ddi_eventcookie_t,
291     void *, void *);
292 static boolean_t xnf_kstat_init(xnf_t *);
293 static void xnf_rx_collect(xnf_t *);
294 
295 #define	XNF_CALLBACK_FLAGS	(MC_GETCAPAB | MC_PROPERTIES)
296 
297 static mac_callbacks_t xnf_callbacks = {
298 	.mc_callbacks = XNF_CALLBACK_FLAGS,
299 	.mc_getstat = xnf_stat,
300 	.mc_start = xnf_start,
301 	.mc_stop = xnf_stop,
302 	.mc_setpromisc = xnf_set_promiscuous,
303 	.mc_multicst = xnf_set_multicast,
304 	.mc_unicst = xnf_set_mac_addr,
305 	.mc_tx = xnf_send,
306 	.mc_getcapab = xnf_getcapab,
307 	.mc_setprop = xnf_setprop,
308 	.mc_getprop = xnf_getprop,
309 	.mc_propinfo = xnf_propinfo,
310 };
311 
312 /* DMA attributes for network ring buffer */
313 static ddi_dma_attr_t ringbuf_dma_attr = {
314 	.dma_attr_version = DMA_ATTR_V0,
315 	.dma_attr_addr_lo = 0,
316 	.dma_attr_addr_hi = 0xffffffffffffffffULL,
317 	.dma_attr_count_max = 0x7fffffff,
318 	.dma_attr_align = MMU_PAGESIZE,
319 	.dma_attr_burstsizes = 0x7ff,
320 	.dma_attr_minxfer = 1,
321 	.dma_attr_maxxfer = 0xffffffffU,
322 	.dma_attr_seg = 0xffffffffffffffffULL,
323 	.dma_attr_sgllen = 1,
324 	.dma_attr_granular = 1,
325 	.dma_attr_flags = 0
326 };
327 
328 /* DMA attributes for receive data */
329 static ddi_dma_attr_t rx_buf_dma_attr = {
330 	.dma_attr_version = DMA_ATTR_V0,
331 	.dma_attr_addr_lo = 0,
332 	.dma_attr_addr_hi = 0xffffffffffffffffULL,
333 	.dma_attr_count_max = MMU_PAGEOFFSET,
334 	.dma_attr_align = MMU_PAGESIZE, /* allocation alignment */
335 	.dma_attr_burstsizes = 0x7ff,
336 	.dma_attr_minxfer = 1,
337 	.dma_attr_maxxfer = 0xffffffffU,
338 	.dma_attr_seg = 0xffffffffffffffffULL,
339 	.dma_attr_sgllen = 1,
340 	.dma_attr_granular = 1,
341 	.dma_attr_flags = 0
342 };
343 
344 /* DMA attributes for transmit data */
345 static ddi_dma_attr_t tx_buf_dma_attr = {
346 	.dma_attr_version = DMA_ATTR_V0,
347 	.dma_attr_addr_lo = 0,
348 	.dma_attr_addr_hi = 0xffffffffffffffffULL,
349 	.dma_attr_count_max = MMU_PAGEOFFSET,
350 	.dma_attr_align = 1,
351 	.dma_attr_burstsizes = 0x7ff,
352 	.dma_attr_minxfer = 1,
353 	.dma_attr_maxxfer = 0xffffffffU,
354 	.dma_attr_seg = XEN_DATA_BOUNDARY - 1, /* segment boundary */
355 	.dma_attr_sgllen = XEN_MAX_TX_DATA_PAGES, /* max number of segments */
356 	.dma_attr_granular = 1,
357 	.dma_attr_flags = 0
358 };
359 
360 /* DMA access attributes for registers and descriptors */
361 static ddi_device_acc_attr_t accattr = {
362 	DDI_DEVICE_ATTR_V0,
363 	DDI_STRUCTURE_LE_ACC,	/* This is a little-endian device */
364 	DDI_STRICTORDER_ACC
365 };
366 
367 /* DMA access attributes for data: NOT to be byte swapped. */
368 static ddi_device_acc_attr_t data_accattr = {
369 	DDI_DEVICE_ATTR_V0,
370 	DDI_NEVERSWAP_ACC,
371 	DDI_STRICTORDER_ACC
372 };
373 
374 DDI_DEFINE_STREAM_OPS(xnf_dev_ops, nulldev, nulldev, xnf_attach, xnf_detach,
375     nodev, NULL, D_MP, NULL, ddi_quiesce_not_supported);
376 
377 static struct modldrv xnf_modldrv = {
378 	&mod_driverops,
379 	"Virtual Ethernet driver",
380 	&xnf_dev_ops
381 };
382 
383 static struct modlinkage modlinkage = {
384 	MODREV_1, &xnf_modldrv, NULL
385 };
386 
387 int
388 _init(void)
389 {
390 	int r;
391 
392 	mac_init_ops(&xnf_dev_ops, "xnf");
393 	r = mod_install(&modlinkage);
394 	if (r != DDI_SUCCESS)
395 		mac_fini_ops(&xnf_dev_ops);
396 
397 	return (r);
398 }
399 
400 int
401 _fini(void)
402 {
403 	return (EBUSY); /* XXPV should be removable */
404 }
405 
406 int
407 _info(struct modinfo *modinfop)
408 {
409 	return (mod_info(&modlinkage, modinfop));
410 }
411 
412 /*
413  * Acquire a grant reference.
414  */
415 static grant_ref_t
416 xnf_gref_get(xnf_t *xnfp)
417 {
418 	grant_ref_t gref;
419 
420 	mutex_enter(&xnfp->xnf_gref_lock);
421 
422 	do {
423 		gref = gnttab_claim_grant_reference(&xnfp->xnf_gref_head);
424 
425 	} while ((gref == INVALID_GRANT_REF) &&
426 	    (gnttab_alloc_grant_references(16, &xnfp->xnf_gref_head) == 0));
427 
428 	mutex_exit(&xnfp->xnf_gref_lock);
429 
430 	if (gref == INVALID_GRANT_REF) {
431 		xnfp->xnf_stat_gref_failure++;
432 	} else {
433 		atomic_inc_64(&xnfp->xnf_stat_gref_outstanding);
434 		if (xnfp->xnf_stat_gref_outstanding > xnfp->xnf_stat_gref_peak)
435 			xnfp->xnf_stat_gref_peak =
436 			    xnfp->xnf_stat_gref_outstanding;
437 	}
438 
439 	return (gref);
440 }
441 
442 /*
443  * Release a grant reference.
444  */
445 static void
446 xnf_gref_put(xnf_t *xnfp, grant_ref_t gref)
447 {
448 	ASSERT(gref != INVALID_GRANT_REF);
449 
450 	mutex_enter(&xnfp->xnf_gref_lock);
451 	gnttab_release_grant_reference(&xnfp->xnf_gref_head, gref);
452 	mutex_exit(&xnfp->xnf_gref_lock);
453 
454 	atomic_dec_64(&xnfp->xnf_stat_gref_outstanding);
455 }
456 
457 /*
458  * Acquire a transmit id.
459  */
460 static xnf_txid_t *
461 xnf_txid_get(xnf_t *xnfp)
462 {
463 	xnf_txid_t *tidp;
464 
465 	ASSERT(MUTEX_HELD(&xnfp->xnf_txlock));
466 
467 	if (xnfp->xnf_tx_pkt_id_head == INVALID_TX_ID)
468 		return (NULL);
469 
470 	ASSERT(TX_ID_VALID(xnfp->xnf_tx_pkt_id_head));
471 
472 	tidp = TX_ID_TO_TXID(xnfp, xnfp->xnf_tx_pkt_id_head);
473 	xnfp->xnf_tx_pkt_id_head = tidp->next;
474 	tidp->next = INVALID_TX_ID;
475 
476 	ASSERT(tidp->txbuf == NULL);
477 
478 	return (tidp);
479 }
480 
481 /*
482  * Release a transmit id.
483  */
484 static void
485 xnf_txid_put(xnf_t *xnfp, xnf_txid_t *tidp)
486 {
487 	ASSERT(MUTEX_HELD(&xnfp->xnf_txlock));
488 	ASSERT(TX_ID_VALID(tidp->id));
489 	ASSERT(tidp->next == INVALID_TX_ID);
490 
491 	tidp->txbuf = NULL;
492 	tidp->next = xnfp->xnf_tx_pkt_id_head;
493 	xnfp->xnf_tx_pkt_id_head = tidp->id;
494 }
495 
496 static void
497 xnf_data_txbuf_free(xnf_t *xnfp, xnf_txbuf_t *txp)
498 {
499 	ASSERT3U(txp->tx_type, ==, TX_DATA);
500 
501 	/*
502 	 * We are either using a lookaside buffer or we are mapping existing
503 	 * buffers.
504 	 */
505 	if (txp->tx_bdesc != NULL) {
506 		ASSERT(!txp->tx_handle_bound);
507 		xnf_buf_put(xnfp, txp->tx_bdesc, B_TRUE);
508 	} else {
509 		if (txp->tx_txreq.gref != INVALID_GRANT_REF) {
510 			if (gnttab_query_foreign_access(txp->tx_txreq.gref) !=
511 			    0) {
512 				cmn_err(CE_PANIC, "tx grant %d still in use by "
513 				    "backend domain", txp->tx_txreq.gref);
514 			}
515 			(void) gnttab_end_foreign_access_ref(
516 			    txp->tx_txreq.gref, 1);
517 			xnf_gref_put(xnfp, txp->tx_txreq.gref);
518 		}
519 
520 		if (txp->tx_handle_bound)
521 			(void) ddi_dma_unbind_handle(txp->tx_dma_handle);
522 	}
523 
524 	if (txp->tx_mp != NULL)
525 		freemsg(txp->tx_mp);
526 
527 	if (txp->tx_prev != NULL) {
528 		ASSERT3P(txp->tx_prev->tx_next, ==, txp);
529 		txp->tx_prev->tx_next = NULL;
530 	}
531 
532 	if (txp->tx_txreq.id != INVALID_TX_ID) {
533 		/*
534 		 * This should be only possible when resuming from a suspend.
535 		 */
536 		ASSERT(!xnfp->xnf_connected);
537 		xnf_txid_put(xnfp, TX_ID_TO_TXID(xnfp, txp->tx_txreq.id));
538 		txp->tx_txreq.id = INVALID_TX_ID;
539 	}
540 
541 	kmem_cache_free(xnfp->xnf_tx_buf_cache, txp);
542 }
543 
544 static void
545 xnf_data_txbuf_free_chain(xnf_t *xnfp, xnf_txbuf_t *txp)
546 {
547 	if (txp == NULL)
548 		return;
549 
550 	while (txp->tx_next != NULL)
551 		txp = txp->tx_next;
552 
553 	/*
554 	 * We free the chain in reverse order so that grants can be released
555 	 * for all dma chunks before unbinding the dma handles. The mblk is
556 	 * freed last, after all its fragments' dma handles are unbound.
557 	 */
558 	xnf_txbuf_t *prev;
559 	for (; txp != NULL; txp = prev) {
560 		prev = txp->tx_prev;
561 		xnf_data_txbuf_free(xnfp, txp);
562 	}
563 }
564 
565 static xnf_txbuf_t *
566 xnf_data_txbuf_alloc(xnf_t *xnfp)
567 {
568 	xnf_txbuf_t *txp = kmem_cache_alloc(xnfp->xnf_tx_buf_cache, KM_SLEEP);
569 	txp->tx_type = TX_DATA;
570 	txp->tx_next = NULL;
571 	txp->tx_prev = NULL;
572 	txp->tx_head = txp;
573 	txp->tx_frags_to_ack = 0;
574 	txp->tx_mp = NULL;
575 	txp->tx_bdesc = NULL;
576 	txp->tx_handle_bound = B_FALSE;
577 	txp->tx_txreq.gref = INVALID_GRANT_REF;
578 	txp->tx_txreq.id = INVALID_TX_ID;
579 
580 	return (txp);
581 }
582 
583 /*
584  * Get `wanted' slots in the transmit ring, waiting for at least that
585  * number if `wait' is B_TRUE. Force the ring to be cleaned by setting
586  * `wanted' to zero.
587  *
588  * Return the number of slots available.
589  */
590 static int
591 xnf_tx_slots_get(xnf_t *xnfp, int wanted, boolean_t wait)
592 {
593 	int slotsfree;
594 	boolean_t forced_clean = (wanted == 0);
595 
596 	ASSERT(MUTEX_HELD(&xnfp->xnf_txlock));
597 
598 	/* LINTED: constant in conditional context */
599 	while (B_TRUE) {
600 		slotsfree = RING_FREE_REQUESTS(&xnfp->xnf_tx_ring);
601 
602 		if ((slotsfree < wanted) || forced_clean)
603 			slotsfree = xnf_tx_clean_ring(xnfp);
604 
605 		/*
606 		 * If there are more than we need free, tell other
607 		 * people to come looking again. We hold txlock, so we
608 		 * are able to take our slots before anyone else runs.
609 		 */
610 		if (slotsfree > wanted)
611 			cv_broadcast(&xnfp->xnf_cv_tx_slots);
612 
613 		if (slotsfree >= wanted)
614 			break;
615 
616 		if (!wait)
617 			break;
618 
619 		cv_wait(&xnfp->xnf_cv_tx_slots, &xnfp->xnf_txlock);
620 	}
621 
622 	ASSERT(slotsfree <= RING_SIZE(&(xnfp->xnf_tx_ring)));
623 
624 	return (slotsfree);
625 }
626 
627 static int
628 xnf_setup_rings(xnf_t *xnfp)
629 {
630 	domid_t			oeid;
631 	struct xenbus_device	*xsd;
632 	RING_IDX		i;
633 	int			err;
634 	xnf_txid_t		*tidp;
635 	xnf_buf_t **bdescp;
636 
637 	oeid = xvdi_get_oeid(xnfp->xnf_devinfo);
638 	xsd = xvdi_get_xsd(xnfp->xnf_devinfo);
639 
640 	if (xnfp->xnf_tx_ring_ref != INVALID_GRANT_REF)
641 		gnttab_end_foreign_access(xnfp->xnf_tx_ring_ref, 0, 0);
642 
643 	err = gnttab_grant_foreign_access(oeid,
644 	    xnf_btop(pa_to_ma(xnfp->xnf_tx_ring_phys_addr)), 0);
645 	if (err <= 0) {
646 		err = -err;
647 		xenbus_dev_error(xsd, err, "granting access to tx ring page");
648 		goto out;
649 	}
650 	xnfp->xnf_tx_ring_ref = (grant_ref_t)err;
651 
652 	if (xnfp->xnf_rx_ring_ref != INVALID_GRANT_REF)
653 		gnttab_end_foreign_access(xnfp->xnf_rx_ring_ref, 0, 0);
654 
655 	err = gnttab_grant_foreign_access(oeid,
656 	    xnf_btop(pa_to_ma(xnfp->xnf_rx_ring_phys_addr)), 0);
657 	if (err <= 0) {
658 		err = -err;
659 		xenbus_dev_error(xsd, err, "granting access to rx ring page");
660 		goto out;
661 	}
662 	xnfp->xnf_rx_ring_ref = (grant_ref_t)err;
663 
664 	mutex_enter(&xnfp->xnf_txlock);
665 
666 	/*
667 	 * We first cleanup the TX ring in case we are doing a resume.
668 	 * Note that this can lose packets, but we expect to stagger on.
669 	 */
670 	xnfp->xnf_tx_pkt_id_head = INVALID_TX_ID; /* I.e. emtpy list. */
671 	for (i = 0, tidp = &xnfp->xnf_tx_pkt_id[0];
672 	    i < NET_TX_RING_SIZE;
673 	    i++, tidp++) {
674 		xnf_txbuf_t *txp = tidp->txbuf;
675 		if (txp == NULL)
676 			continue;
677 
678 		switch (txp->tx_type) {
679 		case TX_DATA:
680 			/*
681 			 * txid_put() will be called for each txbuf's txid in
682 			 * the chain which will result in clearing tidp->txbuf.
683 			 */
684 			xnf_data_txbuf_free_chain(xnfp, txp);
685 
686 			break;
687 
688 		case TX_MCAST_REQ:
689 			txp->tx_type = TX_MCAST_RSP;
690 			txp->tx_status = NETIF_RSP_DROPPED;
691 			cv_broadcast(&xnfp->xnf_cv_multicast);
692 
693 			/*
694 			 * The request consumed two slots in the ring,
695 			 * yet only a single xnf_txid_t is used. Step
696 			 * over the empty slot.
697 			 */
698 			i++;
699 			ASSERT3U(i, <, NET_TX_RING_SIZE);
700 			break;
701 
702 		case TX_MCAST_RSP:
703 			break;
704 		}
705 	}
706 
707 	/*
708 	 * Now purge old list and add each txid to the new free list.
709 	 */
710 	xnfp->xnf_tx_pkt_id_head = INVALID_TX_ID; /* I.e. emtpy list. */
711 	for (i = 0, tidp = &xnfp->xnf_tx_pkt_id[0];
712 	    i < NET_TX_RING_SIZE;
713 	    i++, tidp++) {
714 		tidp->id = i;
715 		ASSERT3P(tidp->txbuf, ==, NULL);
716 		tidp->next = INVALID_TX_ID; /* Appease txid_put(). */
717 		xnf_txid_put(xnfp, tidp);
718 	}
719 
720 	/* LINTED: constant in conditional context */
721 	SHARED_RING_INIT(xnfp->xnf_tx_ring.sring);
722 	/* LINTED: constant in conditional context */
723 	FRONT_RING_INIT(&xnfp->xnf_tx_ring,
724 	    xnfp->xnf_tx_ring.sring, PAGESIZE);
725 
726 	mutex_exit(&xnfp->xnf_txlock);
727 
728 	mutex_enter(&xnfp->xnf_rxlock);
729 
730 	/*
731 	 * Clean out any buffers currently posted to the receive ring
732 	 * before we reset it.
733 	 */
734 	for (i = 0, bdescp = &xnfp->xnf_rx_pkt_info[0];
735 	    i < NET_RX_RING_SIZE;
736 	    i++, bdescp++) {
737 		if (*bdescp != NULL) {
738 			xnf_buf_put(xnfp, *bdescp, B_FALSE);
739 			*bdescp = NULL;
740 		}
741 	}
742 
743 	/* LINTED: constant in conditional context */
744 	SHARED_RING_INIT(xnfp->xnf_rx_ring.sring);
745 	/* LINTED: constant in conditional context */
746 	FRONT_RING_INIT(&xnfp->xnf_rx_ring,
747 	    xnfp->xnf_rx_ring.sring, PAGESIZE);
748 
749 	/*
750 	 * Fill the ring with buffers.
751 	 */
752 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
753 		xnf_buf_t *bdesc;
754 
755 		bdesc = xnf_buf_get(xnfp, KM_SLEEP, B_FALSE);
756 		VERIFY(bdesc != NULL);
757 		xnf_rxbuf_hang(xnfp, bdesc);
758 	}
759 
760 	/* LINTED: constant in conditional context */
761 	RING_PUSH_REQUESTS(&xnfp->xnf_rx_ring);
762 
763 	mutex_exit(&xnfp->xnf_rxlock);
764 
765 	return (0);
766 
767 out:
768 	if (xnfp->xnf_tx_ring_ref != INVALID_GRANT_REF)
769 		gnttab_end_foreign_access(xnfp->xnf_tx_ring_ref, 0, 0);
770 	xnfp->xnf_tx_ring_ref = INVALID_GRANT_REF;
771 
772 	if (xnfp->xnf_rx_ring_ref != INVALID_GRANT_REF)
773 		gnttab_end_foreign_access(xnfp->xnf_rx_ring_ref, 0, 0);
774 	xnfp->xnf_rx_ring_ref = INVALID_GRANT_REF;
775 
776 	return (err);
777 }
778 
779 /*
780  * Connect driver to back end, called to set up communication with
781  * back end driver both initially and on resume after restore/migrate.
782  */
783 void
784 xnf_be_connect(xnf_t *xnfp)
785 {
786 	const char	*message;
787 	xenbus_transaction_t xbt;
788 	struct		xenbus_device *xsd;
789 	char		*xsname;
790 	int		err;
791 
792 	ASSERT(!xnfp->xnf_connected);
793 
794 	xsd = xvdi_get_xsd(xnfp->xnf_devinfo);
795 	xsname = xvdi_get_xsname(xnfp->xnf_devinfo);
796 
797 	err = xnf_setup_rings(xnfp);
798 	if (err != 0) {
799 		cmn_err(CE_WARN, "failed to set up tx/rx rings");
800 		xenbus_dev_error(xsd, err, "setting up ring");
801 		return;
802 	}
803 
804 again:
805 	err = xenbus_transaction_start(&xbt);
806 	if (err != 0) {
807 		xenbus_dev_error(xsd, EIO, "starting transaction");
808 		return;
809 	}
810 
811 	err = xenbus_printf(xbt, xsname, "tx-ring-ref", "%u",
812 	    xnfp->xnf_tx_ring_ref);
813 	if (err != 0) {
814 		message = "writing tx ring-ref";
815 		goto abort_transaction;
816 	}
817 
818 	err = xenbus_printf(xbt, xsname, "rx-ring-ref", "%u",
819 	    xnfp->xnf_rx_ring_ref);
820 	if (err != 0) {
821 		message = "writing rx ring-ref";
822 		goto abort_transaction;
823 	}
824 
825 	err = xenbus_printf(xbt, xsname, "event-channel", "%u",
826 	    xnfp->xnf_evtchn);
827 	if (err != 0) {
828 		message = "writing event-channel";
829 		goto abort_transaction;
830 	}
831 
832 	err = xenbus_printf(xbt, xsname, "feature-rx-notify", "%d", 1);
833 	if (err != 0) {
834 		message = "writing feature-rx-notify";
835 		goto abort_transaction;
836 	}
837 
838 	err = xenbus_printf(xbt, xsname, "request-rx-copy", "%d", 1);
839 	if (err != 0) {
840 		message = "writing request-rx-copy";
841 		goto abort_transaction;
842 	}
843 
844 	if (xnfp->xnf_be_mcast_control) {
845 		err = xenbus_printf(xbt, xsname, "request-multicast-control",
846 		    "%d", 1);
847 		if (err != 0) {
848 			message = "writing request-multicast-control";
849 			goto abort_transaction;
850 		}
851 	}
852 
853 	/*
854 	 * Tell backend if we support scatter-gather lists on the rx side.
855 	 */
856 	err = xenbus_printf(xbt, xsname, "feature-sg", "%d",
857 	    xnf_enable_rx_sg ? 1 : 0);
858 	if (err != 0) {
859 		message = "writing feature-sg";
860 		goto abort_transaction;
861 	}
862 
863 	/*
864 	 * Tell backend if we support LRO for IPv4. Scatter-gather on rx is
865 	 * a prerequisite.
866 	 */
867 	err = xenbus_printf(xbt, xsname, "feature-gso-tcpv4", "%d",
868 	    (xnf_enable_rx_sg && xnf_enable_lro) ? 1 : 0);
869 	if (err != 0) {
870 		message = "writing feature-gso-tcpv4";
871 		goto abort_transaction;
872 	}
873 
874 	err = xvdi_switch_state(xnfp->xnf_devinfo, xbt, XenbusStateConnected);
875 	if (err != 0) {
876 		message = "switching state to XenbusStateConnected";
877 		goto abort_transaction;
878 	}
879 
880 	err = xenbus_transaction_end(xbt, 0);
881 	if (err != 0) {
882 		if (err == EAGAIN)
883 			goto again;
884 		xenbus_dev_error(xsd, err, "completing transaction");
885 	}
886 
887 	return;
888 
889 abort_transaction:
890 	(void) xenbus_transaction_end(xbt, 1);
891 	xenbus_dev_error(xsd, err, "%s", message);
892 }
893 
894 /*
895  * Read configuration information from xenstore.
896  */
897 void
898 xnf_read_config(xnf_t *xnfp)
899 {
900 	int err, be_cap;
901 	char mac[ETHERADDRL * 3];
902 	char *oename = xvdi_get_oename(xnfp->xnf_devinfo);
903 
904 	err = xenbus_scanf(XBT_NULL, oename, "mac",
905 	    "%s", (char *)&mac[0]);
906 	if (err != 0) {
907 		/*
908 		 * bad: we're supposed to be set up with a proper mac
909 		 * addr. at this point
910 		 */
911 		cmn_err(CE_WARN, "%s%d: no mac address",
912 		    ddi_driver_name(xnfp->xnf_devinfo),
913 		    ddi_get_instance(xnfp->xnf_devinfo));
914 			return;
915 	}
916 	if (ether_aton(mac, xnfp->xnf_mac_addr) != ETHERADDRL) {
917 		err = ENOENT;
918 		xenbus_dev_error(xvdi_get_xsd(xnfp->xnf_devinfo), ENOENT,
919 		    "parsing %s/mac", xvdi_get_xsname(xnfp->xnf_devinfo));
920 		return;
921 	}
922 
923 	err = xenbus_scanf(XBT_NULL, oename,
924 	    "feature-rx-copy", "%d", &be_cap);
925 	/*
926 	 * If we fail to read the store we assume that the key is
927 	 * absent, implying an older domain at the far end.  Older
928 	 * domains cannot do HV copy.
929 	 */
930 	if (err != 0)
931 		be_cap = 0;
932 	xnfp->xnf_be_rx_copy = (be_cap != 0);
933 
934 	err = xenbus_scanf(XBT_NULL, oename,
935 	    "feature-multicast-control", "%d", &be_cap);
936 	/*
937 	 * If we fail to read the store we assume that the key is
938 	 * absent, implying an older domain at the far end.  Older
939 	 * domains do not support multicast control.
940 	 */
941 	if (err != 0)
942 		be_cap = 0;
943 	xnfp->xnf_be_mcast_control = (be_cap != 0) && xnf_multicast_control;
944 
945 	/*
946 	 * See if back-end supports scatter-gather for transmits. If not,
947 	 * we will not support LSO and limit the mtu to 1500.
948 	 */
949 	err = xenbus_scanf(XBT_NULL, oename, "feature-sg", "%d", &be_cap);
950 	if (err != 0) {
951 		be_cap = 0;
952 		dev_err(xnfp->xnf_devinfo, CE_WARN, "error reading "
953 		    "'feature-sg' from backend driver");
954 	}
955 	if (be_cap == 0) {
956 		dev_err(xnfp->xnf_devinfo, CE_WARN, "scatter-gather is not "
957 		    "supported for transmits in the backend driver. LSO is "
958 		    "disabled and MTU is restricted to 1500 bytes.");
959 	}
960 	xnfp->xnf_be_tx_sg = (be_cap != 0) && xnf_enable_tx_sg;
961 
962 	if (xnfp->xnf_be_tx_sg) {
963 		/*
964 		 * Check if LSO is supported. Currently we only check for
965 		 * IPv4 as Illumos doesn't support LSO for IPv6.
966 		 */
967 		err = xenbus_scanf(XBT_NULL, oename, "feature-gso-tcpv4", "%d",
968 		    &be_cap);
969 		if (err != 0) {
970 			be_cap = 0;
971 			dev_err(xnfp->xnf_devinfo, CE_WARN, "error reading "
972 			    "'feature-gso-tcpv4' from backend driver");
973 		}
974 		if (be_cap == 0) {
975 			dev_err(xnfp->xnf_devinfo, CE_WARN, "LSO is not "
976 			    "supported by the backend driver. Performance "
977 			    "will be affected.");
978 		}
979 		xnfp->xnf_be_lso = (be_cap != 0) && xnf_enable_lso;
980 	}
981 }
982 
983 /*
984  *  attach(9E) -- Attach a device to the system
985  */
986 static int
987 xnf_attach(dev_info_t *devinfo, ddi_attach_cmd_t cmd)
988 {
989 	mac_register_t *macp;
990 	xnf_t *xnfp;
991 	int err;
992 	char cachename[32];
993 
994 #ifdef XNF_DEBUG
995 	if (xnf_debug & XNF_DEBUG_DDI)
996 		printf("xnf%d: attach(0x%p)\n", ddi_get_instance(devinfo),
997 		    (void *)devinfo);
998 #endif
999 
1000 	switch (cmd) {
1001 	case DDI_RESUME:
1002 		xnfp = ddi_get_driver_private(devinfo);
1003 		xnfp->xnf_gen++;
1004 
1005 		(void) xvdi_resume(devinfo);
1006 		(void) xvdi_alloc_evtchn(devinfo);
1007 		xnfp->xnf_evtchn = xvdi_get_evtchn(devinfo);
1008 #ifdef XPV_HVM_DRIVER
1009 		ec_bind_evtchn_to_handler(xnfp->xnf_evtchn, IPL_VIF, xnf_intr,
1010 		    xnfp);
1011 #else
1012 		(void) ddi_add_intr(devinfo, 0, NULL, NULL, xnf_intr,
1013 		    (caddr_t)xnfp);
1014 #endif
1015 		return (DDI_SUCCESS);
1016 
1017 	case DDI_ATTACH:
1018 		break;
1019 
1020 	default:
1021 		return (DDI_FAILURE);
1022 	}
1023 
1024 	/*
1025 	 *  Allocate gld_mac_info_t and xnf_instance structures
1026 	 */
1027 	macp = mac_alloc(MAC_VERSION);
1028 	if (macp == NULL)
1029 		return (DDI_FAILURE);
1030 	xnfp = kmem_zalloc(sizeof (*xnfp), KM_SLEEP);
1031 
1032 	xnfp->xnf_tx_pkt_id =
1033 	    kmem_zalloc(sizeof (xnf_txid_t) * NET_TX_RING_SIZE, KM_SLEEP);
1034 
1035 	xnfp->xnf_rx_pkt_info =
1036 	    kmem_zalloc(sizeof (xnf_buf_t *) * NET_RX_RING_SIZE, KM_SLEEP);
1037 
1038 	macp->m_dip = devinfo;
1039 	macp->m_driver = xnfp;
1040 	xnfp->xnf_devinfo = devinfo;
1041 
1042 	macp->m_type_ident = MAC_PLUGIN_IDENT_ETHER;
1043 	macp->m_src_addr = xnfp->xnf_mac_addr;
1044 	macp->m_callbacks = &xnf_callbacks;
1045 	macp->m_min_sdu = 0;
1046 	xnfp->xnf_mtu = ETHERMTU;
1047 	macp->m_max_sdu = xnfp->xnf_mtu;
1048 
1049 	xnfp->xnf_running = B_FALSE;
1050 	xnfp->xnf_connected = B_FALSE;
1051 	xnfp->xnf_be_rx_copy = B_FALSE;
1052 	xnfp->xnf_be_mcast_control = B_FALSE;
1053 	xnfp->xnf_need_sched = B_FALSE;
1054 
1055 	xnfp->xnf_rx_head = NULL;
1056 	xnfp->xnf_rx_tail = NULL;
1057 	xnfp->xnf_rx_new_buffers_posted = B_FALSE;
1058 
1059 #ifdef XPV_HVM_DRIVER
1060 	/* Report our version to dom0 */
1061 	(void) xenbus_printf(XBT_NULL, "guest/xnf", "version", "%d",
1062 	    HVMPV_XNF_VERS);
1063 #endif
1064 
1065 	/*
1066 	 * Get the iblock cookie with which to initialize the mutexes.
1067 	 */
1068 	if (ddi_get_iblock_cookie(devinfo, 0, &xnfp->xnf_icookie)
1069 	    != DDI_SUCCESS)
1070 		goto failure;
1071 
1072 	mutex_init(&xnfp->xnf_txlock,
1073 	    NULL, MUTEX_DRIVER, xnfp->xnf_icookie);
1074 	mutex_init(&xnfp->xnf_rxlock,
1075 	    NULL, MUTEX_DRIVER, xnfp->xnf_icookie);
1076 	mutex_init(&xnfp->xnf_schedlock,
1077 	    NULL, MUTEX_DRIVER, xnfp->xnf_icookie);
1078 	mutex_init(&xnfp->xnf_gref_lock,
1079 	    NULL, MUTEX_DRIVER, xnfp->xnf_icookie);
1080 
1081 	cv_init(&xnfp->xnf_cv_state, NULL, CV_DEFAULT, NULL);
1082 	cv_init(&xnfp->xnf_cv_multicast, NULL, CV_DEFAULT, NULL);
1083 	cv_init(&xnfp->xnf_cv_tx_slots, NULL, CV_DEFAULT, NULL);
1084 
1085 	(void) sprintf(cachename, "xnf_buf_cache_%d",
1086 	    ddi_get_instance(devinfo));
1087 	xnfp->xnf_buf_cache = kmem_cache_create(cachename,
1088 	    sizeof (xnf_buf_t), 0,
1089 	    xnf_buf_constructor, xnf_buf_destructor,
1090 	    NULL, xnfp, NULL, 0);
1091 	if (xnfp->xnf_buf_cache == NULL)
1092 		goto failure_0;
1093 
1094 	(void) sprintf(cachename, "xnf_tx_buf_cache_%d",
1095 	    ddi_get_instance(devinfo));
1096 	xnfp->xnf_tx_buf_cache = kmem_cache_create(cachename,
1097 	    sizeof (xnf_txbuf_t), 0,
1098 	    xnf_tx_buf_constructor, xnf_tx_buf_destructor,
1099 	    NULL, xnfp, NULL, 0);
1100 	if (xnfp->xnf_tx_buf_cache == NULL)
1101 		goto failure_1;
1102 
1103 	xnfp->xnf_gref_head = INVALID_GRANT_REF;
1104 
1105 	if (xnf_alloc_dma_resources(xnfp) == DDI_FAILURE) {
1106 		cmn_err(CE_WARN, "xnf%d: failed to allocate and initialize "
1107 		    "driver data structures",
1108 		    ddi_get_instance(xnfp->xnf_devinfo));
1109 		goto failure_2;
1110 	}
1111 
1112 	xnfp->xnf_rx_ring.sring->rsp_event =
1113 	    xnfp->xnf_tx_ring.sring->rsp_event = 1;
1114 
1115 	xnfp->xnf_tx_ring_ref = INVALID_GRANT_REF;
1116 	xnfp->xnf_rx_ring_ref = INVALID_GRANT_REF;
1117 
1118 	/* set driver private pointer now */
1119 	ddi_set_driver_private(devinfo, xnfp);
1120 
1121 	if (!xnf_kstat_init(xnfp))
1122 		goto failure_3;
1123 
1124 	/*
1125 	 * Allocate an event channel, add the interrupt handler and
1126 	 * bind it to the event channel.
1127 	 */
1128 	(void) xvdi_alloc_evtchn(devinfo);
1129 	xnfp->xnf_evtchn = xvdi_get_evtchn(devinfo);
1130 #ifdef XPV_HVM_DRIVER
1131 	ec_bind_evtchn_to_handler(xnfp->xnf_evtchn, IPL_VIF, xnf_intr, xnfp);
1132 #else
1133 	(void) ddi_add_intr(devinfo, 0, NULL, NULL, xnf_intr, (caddr_t)xnfp);
1134 #endif
1135 
1136 	err = mac_register(macp, &xnfp->xnf_mh);
1137 	mac_free(macp);
1138 	macp = NULL;
1139 	if (err != 0)
1140 		goto failure_4;
1141 
1142 	if (xvdi_add_event_handler(devinfo, XS_OE_STATE, oe_state_change, NULL)
1143 	    != DDI_SUCCESS)
1144 		goto failure_5;
1145 
1146 #ifdef XPV_HVM_DRIVER
1147 	/*
1148 	 * In the HVM case, this driver essentially replaces a driver for
1149 	 * a 'real' PCI NIC. Without the "model" property set to
1150 	 * "Ethernet controller", like the PCI code does, netbooting does
1151 	 * not work correctly, as strplumb_get_netdev_path() will not find
1152 	 * this interface.
1153 	 */
1154 	(void) ndi_prop_update_string(DDI_DEV_T_NONE, devinfo, "model",
1155 	    "Ethernet controller");
1156 #endif
1157 
1158 #ifdef XNF_DEBUG
1159 	if (xnf_debug_instance == NULL)
1160 		xnf_debug_instance = xnfp;
1161 #endif
1162 
1163 	return (DDI_SUCCESS);
1164 
1165 failure_5:
1166 	(void) mac_unregister(xnfp->xnf_mh);
1167 
1168 failure_4:
1169 #ifdef XPV_HVM_DRIVER
1170 	ec_unbind_evtchn(xnfp->xnf_evtchn);
1171 	xvdi_free_evtchn(devinfo);
1172 #else
1173 	ddi_remove_intr(devinfo, 0, xnfp->xnf_icookie);
1174 #endif
1175 	xnfp->xnf_evtchn = INVALID_EVTCHN;
1176 	kstat_delete(xnfp->xnf_kstat_aux);
1177 
1178 failure_3:
1179 	xnf_release_dma_resources(xnfp);
1180 
1181 failure_2:
1182 	kmem_cache_destroy(xnfp->xnf_tx_buf_cache);
1183 
1184 failure_1:
1185 	kmem_cache_destroy(xnfp->xnf_buf_cache);
1186 
1187 failure_0:
1188 	cv_destroy(&xnfp->xnf_cv_tx_slots);
1189 	cv_destroy(&xnfp->xnf_cv_multicast);
1190 	cv_destroy(&xnfp->xnf_cv_state);
1191 
1192 	mutex_destroy(&xnfp->xnf_gref_lock);
1193 	mutex_destroy(&xnfp->xnf_schedlock);
1194 	mutex_destroy(&xnfp->xnf_rxlock);
1195 	mutex_destroy(&xnfp->xnf_txlock);
1196 
1197 failure:
1198 	kmem_free(xnfp, sizeof (*xnfp));
1199 	if (macp != NULL)
1200 		mac_free(macp);
1201 
1202 	return (DDI_FAILURE);
1203 }
1204 
1205 /*  detach(9E) -- Detach a device from the system */
1206 static int
1207 xnf_detach(dev_info_t *devinfo, ddi_detach_cmd_t cmd)
1208 {
1209 	xnf_t *xnfp;		/* Our private device info */
1210 
1211 #ifdef XNF_DEBUG
1212 	if (xnf_debug & XNF_DEBUG_DDI)
1213 		printf("xnf_detach(0x%p)\n", (void *)devinfo);
1214 #endif
1215 
1216 	xnfp = ddi_get_driver_private(devinfo);
1217 
1218 	switch (cmd) {
1219 	case DDI_SUSPEND:
1220 #ifdef XPV_HVM_DRIVER
1221 		ec_unbind_evtchn(xnfp->xnf_evtchn);
1222 		xvdi_free_evtchn(devinfo);
1223 #else
1224 		ddi_remove_intr(devinfo, 0, xnfp->xnf_icookie);
1225 #endif
1226 
1227 		xvdi_suspend(devinfo);
1228 
1229 		mutex_enter(&xnfp->xnf_rxlock);
1230 		mutex_enter(&xnfp->xnf_txlock);
1231 
1232 		xnfp->xnf_evtchn = INVALID_EVTCHN;
1233 		xnfp->xnf_connected = B_FALSE;
1234 		mutex_exit(&xnfp->xnf_txlock);
1235 		mutex_exit(&xnfp->xnf_rxlock);
1236 
1237 		/* claim link to be down after disconnect */
1238 		mac_link_update(xnfp->xnf_mh, LINK_STATE_DOWN);
1239 		return (DDI_SUCCESS);
1240 
1241 	case DDI_DETACH:
1242 		break;
1243 
1244 	default:
1245 		return (DDI_FAILURE);
1246 	}
1247 
1248 	if (xnfp->xnf_connected)
1249 		return (DDI_FAILURE);
1250 
1251 	/*
1252 	 * Cannot detach if we have xnf_buf_t outstanding.
1253 	 */
1254 	if (xnfp->xnf_stat_buf_allocated > 0)
1255 		return (DDI_FAILURE);
1256 
1257 	if (mac_unregister(xnfp->xnf_mh) != 0)
1258 		return (DDI_FAILURE);
1259 
1260 	kstat_delete(xnfp->xnf_kstat_aux);
1261 
1262 	/* Stop the receiver */
1263 	xnf_stop(xnfp);
1264 
1265 	xvdi_remove_event_handler(devinfo, XS_OE_STATE);
1266 
1267 	/* Remove the interrupt */
1268 #ifdef XPV_HVM_DRIVER
1269 	ec_unbind_evtchn(xnfp->xnf_evtchn);
1270 	xvdi_free_evtchn(devinfo);
1271 #else
1272 	ddi_remove_intr(devinfo, 0, xnfp->xnf_icookie);
1273 #endif
1274 
1275 	/* Release any pending xmit mblks */
1276 	xnf_release_mblks(xnfp);
1277 
1278 	/* Release all DMA resources */
1279 	xnf_release_dma_resources(xnfp);
1280 
1281 	cv_destroy(&xnfp->xnf_cv_tx_slots);
1282 	cv_destroy(&xnfp->xnf_cv_multicast);
1283 	cv_destroy(&xnfp->xnf_cv_state);
1284 
1285 	kmem_cache_destroy(xnfp->xnf_tx_buf_cache);
1286 	kmem_cache_destroy(xnfp->xnf_buf_cache);
1287 
1288 	mutex_destroy(&xnfp->xnf_gref_lock);
1289 	mutex_destroy(&xnfp->xnf_schedlock);
1290 	mutex_destroy(&xnfp->xnf_rxlock);
1291 	mutex_destroy(&xnfp->xnf_txlock);
1292 
1293 	kmem_free(xnfp, sizeof (*xnfp));
1294 
1295 	return (DDI_SUCCESS);
1296 }
1297 
1298 /*
1299  *  xnf_set_mac_addr() -- set the physical network address on the board.
1300  */
1301 static int
1302 xnf_set_mac_addr(void *arg, const uint8_t *macaddr)
1303 {
1304 	_NOTE(ARGUNUSED(arg, macaddr));
1305 
1306 	/*
1307 	 * We can't set our macaddr.
1308 	 */
1309 	return (ENOTSUP);
1310 }
1311 
1312 /*
1313  *  xnf_set_multicast() -- set (enable) or disable a multicast address.
1314  *
1315  *  Program the hardware to enable/disable the multicast address
1316  *  in "mca".  Enable if "add" is true, disable if false.
1317  */
1318 static int
1319 xnf_set_multicast(void *arg, boolean_t add, const uint8_t *mca)
1320 {
1321 	xnf_t *xnfp = arg;
1322 	xnf_txbuf_t *txp;
1323 	int n_slots;
1324 	RING_IDX slot;
1325 	xnf_txid_t *tidp;
1326 	netif_tx_request_t *txrp;
1327 	struct netif_extra_info *erp;
1328 	boolean_t notify, result;
1329 
1330 	/*
1331 	 * If the backend does not support multicast control then we
1332 	 * must assume that the right packets will just arrive.
1333 	 */
1334 	if (!xnfp->xnf_be_mcast_control)
1335 		return (0);
1336 
1337 	txp = kmem_cache_alloc(xnfp->xnf_tx_buf_cache, KM_SLEEP);
1338 
1339 	mutex_enter(&xnfp->xnf_txlock);
1340 
1341 	/*
1342 	 * If we're not yet connected then claim success. This is
1343 	 * acceptable because we refresh the entire set of multicast
1344 	 * addresses when we get connected.
1345 	 *
1346 	 * We can't wait around here because the MAC layer expects
1347 	 * this to be a non-blocking operation - waiting ends up
1348 	 * causing a deadlock during resume.
1349 	 */
1350 	if (!xnfp->xnf_connected) {
1351 		mutex_exit(&xnfp->xnf_txlock);
1352 		return (0);
1353 	}
1354 
1355 	/*
1356 	 * 1. Acquire two slots in the ring.
1357 	 * 2. Fill in the slots.
1358 	 * 3. Request notification when the operation is done.
1359 	 * 4. Kick the peer.
1360 	 * 5. Wait for the response via xnf_tx_clean_ring().
1361 	 */
1362 
1363 	n_slots = xnf_tx_slots_get(xnfp, 2, B_TRUE);
1364 	ASSERT(n_slots >= 2);
1365 
1366 	slot = xnfp->xnf_tx_ring.req_prod_pvt;
1367 	tidp = xnf_txid_get(xnfp);
1368 	VERIFY(tidp != NULL);
1369 
1370 	txp->tx_type = TX_MCAST_REQ;
1371 	txp->tx_slot = slot;
1372 
1373 	txrp = RING_GET_REQUEST(&xnfp->xnf_tx_ring, slot);
1374 	erp = (struct netif_extra_info *)
1375 	    RING_GET_REQUEST(&xnfp->xnf_tx_ring, slot + 1);
1376 
1377 	txrp->gref = 0;
1378 	txrp->size = 0;
1379 	txrp->offset = 0;
1380 	/* Set tx_txreq.id to appease xnf_tx_clean_ring(). */
1381 	txrp->id = txp->tx_txreq.id = tidp->id;
1382 	txrp->flags = NETTXF_extra_info;
1383 
1384 	erp->type = add ? XEN_NETIF_EXTRA_TYPE_MCAST_ADD :
1385 	    XEN_NETIF_EXTRA_TYPE_MCAST_DEL;
1386 	bcopy((void *)mca, &erp->u.mcast.addr, ETHERADDRL);
1387 
1388 	tidp->txbuf = txp;
1389 
1390 	xnfp->xnf_tx_ring.req_prod_pvt = slot + 2;
1391 
1392 	mutex_enter(&xnfp->xnf_schedlock);
1393 	xnfp->xnf_pending_multicast++;
1394 	mutex_exit(&xnfp->xnf_schedlock);
1395 
1396 	/* LINTED: constant in conditional context */
1397 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&xnfp->xnf_tx_ring,
1398 	    notify);
1399 	if (notify)
1400 		ec_notify_via_evtchn(xnfp->xnf_evtchn);
1401 
1402 	while (txp->tx_type == TX_MCAST_REQ)
1403 		cv_wait(&xnfp->xnf_cv_multicast, &xnfp->xnf_txlock);
1404 
1405 	ASSERT3U(txp->tx_type, ==, TX_MCAST_RSP);
1406 
1407 	mutex_enter(&xnfp->xnf_schedlock);
1408 	xnfp->xnf_pending_multicast--;
1409 	mutex_exit(&xnfp->xnf_schedlock);
1410 
1411 	result = (txp->tx_status == NETIF_RSP_OKAY);
1412 
1413 	xnf_txid_put(xnfp, tidp);
1414 
1415 	mutex_exit(&xnfp->xnf_txlock);
1416 
1417 	kmem_cache_free(xnfp->xnf_tx_buf_cache, txp);
1418 
1419 	return (result ? 0 : 1);
1420 }
1421 
1422 /*
1423  * xnf_set_promiscuous() -- set or reset promiscuous mode on the board
1424  *
1425  *  Program the hardware to enable/disable promiscuous mode.
1426  */
1427 static int
1428 xnf_set_promiscuous(void *arg, boolean_t on)
1429 {
1430 	_NOTE(ARGUNUSED(arg, on));
1431 
1432 	/*
1433 	 * We can't really do this, but we pretend that we can in
1434 	 * order that snoop will work.
1435 	 */
1436 	return (0);
1437 }
1438 
1439 /*
1440  * Clean buffers that we have responses for from the transmit ring.
1441  */
1442 static int
1443 xnf_tx_clean_ring(xnf_t *xnfp)
1444 {
1445 	boolean_t work_to_do;
1446 
1447 	ASSERT(MUTEX_HELD(&xnfp->xnf_txlock));
1448 
1449 loop:
1450 	while (RING_HAS_UNCONSUMED_RESPONSES(&xnfp->xnf_tx_ring)) {
1451 		RING_IDX cons, prod, i;
1452 
1453 		cons = xnfp->xnf_tx_ring.rsp_cons;
1454 		prod = xnfp->xnf_tx_ring.sring->rsp_prod;
1455 		membar_consumer();
1456 		/*
1457 		 * Clean tx requests from ring that we have responses
1458 		 * for.
1459 		 */
1460 		DTRACE_PROBE2(xnf_tx_clean_range, int, cons, int, prod);
1461 		for (i = cons; i != prod; i++) {
1462 			netif_tx_response_t *trp;
1463 			xnf_txid_t *tidp;
1464 			xnf_txbuf_t *txp;
1465 
1466 			trp = RING_GET_RESPONSE(&xnfp->xnf_tx_ring, i);
1467 			/*
1468 			 * if this slot was occupied by netif_extra_info_t,
1469 			 * then the response will be NETIF_RSP_NULL. In this
1470 			 * case there are no resources to clean up.
1471 			 */
1472 			if (trp->status == NETIF_RSP_NULL)
1473 				continue;
1474 
1475 			ASSERT(TX_ID_VALID(trp->id));
1476 
1477 			tidp = TX_ID_TO_TXID(xnfp, trp->id);
1478 			ASSERT3U(tidp->id, ==, trp->id);
1479 			ASSERT3U(tidp->next, ==, INVALID_TX_ID);
1480 
1481 			txp = tidp->txbuf;
1482 			ASSERT(txp != NULL);
1483 			ASSERT3U(txp->tx_txreq.id, ==, trp->id);
1484 
1485 			switch (txp->tx_type) {
1486 			case TX_DATA:
1487 				/*
1488 				 * We must put the txid for each response we
1489 				 * acknowledge to make sure that we never have
1490 				 * more free slots than txids. Because of this
1491 				 * we do it here instead of waiting for it to
1492 				 * be done in xnf_data_txbuf_free_chain().
1493 				 */
1494 				xnf_txid_put(xnfp, tidp);
1495 				txp->tx_txreq.id = INVALID_TX_ID;
1496 				ASSERT3S(txp->tx_head->tx_frags_to_ack, >, 0);
1497 				txp->tx_head->tx_frags_to_ack--;
1498 
1499 				/*
1500 				 * We clean the whole chain once we got a
1501 				 * response for each fragment.
1502 				 */
1503 				if (txp->tx_head->tx_frags_to_ack == 0)
1504 					xnf_data_txbuf_free_chain(xnfp, txp);
1505 
1506 				break;
1507 
1508 			case TX_MCAST_REQ:
1509 				txp->tx_type = TX_MCAST_RSP;
1510 				txp->tx_status = trp->status;
1511 				cv_broadcast(&xnfp->xnf_cv_multicast);
1512 
1513 				break;
1514 
1515 			default:
1516 				cmn_err(CE_PANIC, "xnf_tx_clean_ring: "
1517 				    "invalid xnf_txbuf_t type: %d",
1518 				    txp->tx_type);
1519 				break;
1520 			}
1521 		}
1522 		/*
1523 		 * Record the last response we dealt with so that we
1524 		 * know where to start next time around.
1525 		 */
1526 		xnfp->xnf_tx_ring.rsp_cons = prod;
1527 		membar_enter();
1528 	}
1529 
1530 	/* LINTED: constant in conditional context */
1531 	RING_FINAL_CHECK_FOR_RESPONSES(&xnfp->xnf_tx_ring, work_to_do);
1532 	if (work_to_do)
1533 		goto loop;
1534 
1535 	return (RING_FREE_REQUESTS(&xnfp->xnf_tx_ring));
1536 }
1537 
1538 /*
1539  * Allocate and fill in a look-aside buffer for the packet `mp'. Used
1540  * to ensure that the packet is physically contiguous and contained
1541  * within a single page.
1542  */
1543 static xnf_buf_t *
1544 xnf_tx_get_lookaside(xnf_t *xnfp, mblk_t *mp, size_t *plen)
1545 {
1546 	xnf_buf_t *bd;
1547 	caddr_t bp;
1548 
1549 	bd = xnf_buf_get(xnfp, KM_SLEEP, B_TRUE);
1550 	if (bd == NULL)
1551 		return (NULL);
1552 
1553 	bp = bd->buf;
1554 	while (mp != NULL) {
1555 		size_t len = MBLKL(mp);
1556 
1557 		bcopy(mp->b_rptr, bp, len);
1558 		bp += len;
1559 
1560 		mp = mp->b_cont;
1561 	}
1562 
1563 	*plen = bp - bd->buf;
1564 	ASSERT3U(*plen, <=, PAGESIZE);
1565 
1566 	xnfp->xnf_stat_tx_lookaside++;
1567 
1568 	return (bd);
1569 }
1570 
1571 /*
1572  * Insert the pseudo-header checksum into the packet.
1573  * Assumes packet is IPv4, TCP/UDP since we only advertised support for
1574  * HCKSUM_INET_FULL_V4.
1575  */
1576 int
1577 xnf_pseudo_cksum(mblk_t *mp)
1578 {
1579 	struct ether_header *ehp;
1580 	uint16_t sap, iplen, *stuff;
1581 	uint32_t cksum;
1582 	size_t len;
1583 	ipha_t *ipha;
1584 	ipaddr_t src, dst;
1585 	uchar_t *ptr;
1586 
1587 	ptr = mp->b_rptr;
1588 	len = MBLKL(mp);
1589 
1590 	/* Each header must fit completely in an mblk. */
1591 	ASSERT3U(len, >=, sizeof (*ehp));
1592 
1593 	ehp = (struct ether_header *)ptr;
1594 
1595 	if (ntohs(ehp->ether_type) == VLAN_TPID) {
1596 		struct ether_vlan_header *evhp;
1597 		ASSERT3U(len, >=, sizeof (*evhp));
1598 		evhp = (struct ether_vlan_header *)ptr;
1599 		sap = ntohs(evhp->ether_type);
1600 		ptr += sizeof (*evhp);
1601 		len -= sizeof (*evhp);
1602 	} else {
1603 		sap = ntohs(ehp->ether_type);
1604 		ptr += sizeof (*ehp);
1605 		len -= sizeof (*ehp);
1606 	}
1607 
1608 	ASSERT3U(sap, ==, ETHERTYPE_IP);
1609 
1610 	/*
1611 	 * Ethernet and IP headers may be in different mblks.
1612 	 */
1613 	ASSERT3P(ptr, <=, mp->b_wptr);
1614 	if (ptr == mp->b_wptr) {
1615 		mp = mp->b_cont;
1616 		ptr = mp->b_rptr;
1617 		len = MBLKL(mp);
1618 	}
1619 
1620 	ASSERT3U(len, >=, sizeof (ipha_t));
1621 	ipha = (ipha_t *)ptr;
1622 
1623 	/*
1624 	 * We assume the IP header has no options. (This is enforced in
1625 	 * ire_send_wire_v4() -- search for IXAF_NO_HW_CKSUM).
1626 	 */
1627 	ASSERT3U(IPH_HDR_LENGTH(ipha), ==, IP_SIMPLE_HDR_LENGTH);
1628 	iplen = ntohs(ipha->ipha_length) - IP_SIMPLE_HDR_LENGTH;
1629 
1630 	ptr += IP_SIMPLE_HDR_LENGTH;
1631 	len -= IP_SIMPLE_HDR_LENGTH;
1632 
1633 	/*
1634 	 * IP and L4 headers may be in different mblks.
1635 	 */
1636 	ASSERT3P(ptr, <=, mp->b_wptr);
1637 	if (ptr == mp->b_wptr) {
1638 		mp = mp->b_cont;
1639 		ptr = mp->b_rptr;
1640 		len = MBLKL(mp);
1641 	}
1642 
1643 	switch (ipha->ipha_protocol) {
1644 	case IPPROTO_TCP:
1645 		ASSERT3U(len, >=, sizeof (tcph_t));
1646 		stuff = (uint16_t *)(ptr + TCP_CHECKSUM_OFFSET);
1647 		cksum = IP_TCP_CSUM_COMP;
1648 		break;
1649 	case IPPROTO_UDP:
1650 		ASSERT3U(len, >=, sizeof (struct udphdr));
1651 		stuff = (uint16_t *)(ptr + UDP_CHECKSUM_OFFSET);
1652 		cksum = IP_UDP_CSUM_COMP;
1653 		break;
1654 	default:
1655 		cmn_err(CE_WARN, "xnf_pseudo_cksum: unexpected protocol %d",
1656 		    ipha->ipha_protocol);
1657 		return (EINVAL);
1658 	}
1659 
1660 	src = ipha->ipha_src;
1661 	dst = ipha->ipha_dst;
1662 
1663 	cksum += (dst >> 16) + (dst & 0xFFFF);
1664 	cksum += (src >> 16) + (src & 0xFFFF);
1665 	cksum += htons(iplen);
1666 
1667 	cksum = (cksum >> 16) + (cksum & 0xFFFF);
1668 	cksum = (cksum >> 16) + (cksum & 0xFFFF);
1669 
1670 	ASSERT(cksum <= 0xFFFF);
1671 
1672 	*stuff = (uint16_t)(cksum ? cksum : ~cksum);
1673 
1674 	return (0);
1675 }
1676 
1677 /*
1678  * Push a packet into the transmit ring.
1679  *
1680  * Note: the format of a tx packet that spans multiple slots is similar to
1681  * what is described in xnf_rx_one_packet().
1682  */
1683 static void
1684 xnf_tx_push_packet(xnf_t *xnfp, xnf_txbuf_t *head)
1685 {
1686 	int nslots = 0;
1687 	int extras = 0;
1688 	RING_IDX slot;
1689 	boolean_t notify;
1690 
1691 	ASSERT(MUTEX_HELD(&xnfp->xnf_txlock));
1692 	ASSERT(xnfp->xnf_running);
1693 
1694 	slot = xnfp->xnf_tx_ring.req_prod_pvt;
1695 
1696 	/*
1697 	 * The caller has already checked that we have enough slots to proceed.
1698 	 */
1699 	for (xnf_txbuf_t *txp = head; txp != NULL; txp = txp->tx_next) {
1700 		xnf_txid_t *tidp;
1701 		netif_tx_request_t *txrp;
1702 
1703 		tidp = xnf_txid_get(xnfp);
1704 		VERIFY(tidp != NULL);
1705 		txrp = RING_GET_REQUEST(&xnfp->xnf_tx_ring, slot);
1706 
1707 		txp->tx_slot = slot;
1708 		txp->tx_txreq.id = tidp->id;
1709 		*txrp = txp->tx_txreq;
1710 
1711 		tidp->txbuf = txp;
1712 		slot++;
1713 		nslots++;
1714 
1715 		/*
1716 		 * When present, LSO info is placed in a slot after the first
1717 		 * data segment, and doesn't require a txid.
1718 		 */
1719 		if (txp->tx_txreq.flags & NETTXF_extra_info) {
1720 			netif_extra_info_t *extra;
1721 			ASSERT3U(nslots, ==, 1);
1722 
1723 			extra = (netif_extra_info_t *)
1724 			    RING_GET_REQUEST(&xnfp->xnf_tx_ring, slot);
1725 			*extra = txp->tx_extra;
1726 			slot++;
1727 			nslots++;
1728 			extras = 1;
1729 		}
1730 	}
1731 
1732 	ASSERT3U(nslots, <=, XEN_MAX_SLOTS_PER_TX);
1733 
1734 	/*
1735 	 * Store the number of data fragments.
1736 	 */
1737 	head->tx_frags_to_ack = nslots - extras;
1738 
1739 	xnfp->xnf_tx_ring.req_prod_pvt = slot;
1740 
1741 	/*
1742 	 * Tell the peer that we sent something, if it cares.
1743 	 */
1744 	/* LINTED: constant in conditional context */
1745 	RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&xnfp->xnf_tx_ring, notify);
1746 	if (notify)
1747 		ec_notify_via_evtchn(xnfp->xnf_evtchn);
1748 }
1749 
1750 static xnf_txbuf_t *
1751 xnf_mblk_copy(xnf_t *xnfp, mblk_t *mp)
1752 {
1753 	xnf_txbuf_t *txp = xnf_data_txbuf_alloc(xnfp);
1754 	size_t length;
1755 
1756 	txp->tx_bdesc = xnf_tx_get_lookaside(xnfp, mp, &length);
1757 	if (txp->tx_bdesc == NULL) {
1758 		xnf_data_txbuf_free(xnfp, txp);
1759 		return (NULL);
1760 	}
1761 	txp->tx_mfn = txp->tx_bdesc->buf_mfn;
1762 	txp->tx_txreq.gref = txp->tx_bdesc->grant_ref;
1763 	txp->tx_txreq.size = length;
1764 	txp->tx_txreq.offset = (uintptr_t)txp->tx_bdesc->buf & PAGEOFFSET;
1765 	txp->tx_txreq.flags = 0;
1766 
1767 	return (txp);
1768 }
1769 
1770 static xnf_txbuf_t *
1771 xnf_mblk_map(xnf_t *xnfp, mblk_t *mp, int *countp)
1772 {
1773 	xnf_txbuf_t *head = NULL;
1774 	xnf_txbuf_t *tail = NULL;
1775 	domid_t oeid;
1776 	int nsegs = 0;
1777 
1778 	oeid = xvdi_get_oeid(xnfp->xnf_devinfo);
1779 
1780 	for (mblk_t *ml = mp; ml != NULL; ml = ml->b_cont) {
1781 		ddi_dma_handle_t dma_handle;
1782 		ddi_dma_cookie_t dma_cookie;
1783 		uint_t ncookies;
1784 		xnf_txbuf_t *txp;
1785 
1786 		if (MBLKL(ml) == 0)
1787 			continue;
1788 
1789 		txp = xnf_data_txbuf_alloc(xnfp);
1790 
1791 		if (head == NULL) {
1792 			head = txp;
1793 		} else {
1794 			ASSERT(tail != NULL);
1795 			TXBUF_SETNEXT(tail, txp);
1796 			txp->tx_head = head;
1797 		}
1798 
1799 		/*
1800 		 * The necessary segmentation rules (e.g. not crossing a page
1801 		 * boundary) are enforced by the dma attributes of the handle.
1802 		 */
1803 		dma_handle = txp->tx_dma_handle;
1804 		int ret = ddi_dma_addr_bind_handle(dma_handle,
1805 		    NULL, (char *)ml->b_rptr, MBLKL(ml),
1806 		    DDI_DMA_WRITE | DDI_DMA_STREAMING,
1807 		    DDI_DMA_DONTWAIT, 0, &dma_cookie,
1808 		    &ncookies);
1809 		if (ret != DDI_DMA_MAPPED) {
1810 			if (ret != DDI_DMA_NORESOURCES) {
1811 				dev_err(xnfp->xnf_devinfo, CE_WARN,
1812 				    "ddi_dma_addr_bind_handle() failed "
1813 				    "[dma_error=%d]", ret);
1814 			}
1815 			goto error;
1816 		}
1817 		txp->tx_handle_bound = B_TRUE;
1818 
1819 		ASSERT(ncookies > 0);
1820 		for (int i = 0; i < ncookies; i++) {
1821 			if (nsegs == XEN_MAX_TX_DATA_PAGES) {
1822 				dev_err(xnfp->xnf_devinfo, CE_WARN,
1823 				    "xnf_dmamap_alloc() failed: "
1824 				    "too many segments");
1825 				goto error;
1826 			}
1827 			if (i > 0) {
1828 				txp = xnf_data_txbuf_alloc(xnfp);
1829 				ASSERT(tail != NULL);
1830 				TXBUF_SETNEXT(tail, txp);
1831 				txp->tx_head = head;
1832 			}
1833 
1834 			txp->tx_mfn =
1835 			    xnf_btop(pa_to_ma(dma_cookie.dmac_laddress));
1836 			txp->tx_txreq.gref = xnf_gref_get(xnfp);
1837 			if (txp->tx_txreq.gref == INVALID_GRANT_REF) {
1838 				dev_err(xnfp->xnf_devinfo, CE_WARN,
1839 				    "xnf_dmamap_alloc() failed: "
1840 				    "invalid grant ref");
1841 				goto error;
1842 			}
1843 			gnttab_grant_foreign_access_ref(txp->tx_txreq.gref,
1844 			    oeid, txp->tx_mfn, 1);
1845 			txp->tx_txreq.offset =
1846 			    dma_cookie.dmac_laddress & PAGEOFFSET;
1847 			txp->tx_txreq.size = dma_cookie.dmac_size;
1848 			txp->tx_txreq.flags = 0;
1849 
1850 			ddi_dma_nextcookie(dma_handle, &dma_cookie);
1851 			nsegs++;
1852 
1853 			if (tail != NULL)
1854 				tail->tx_txreq.flags = NETTXF_more_data;
1855 			tail = txp;
1856 		}
1857 	}
1858 
1859 	*countp = nsegs;
1860 	return (head);
1861 
1862 error:
1863 	xnf_data_txbuf_free_chain(xnfp, head);
1864 	return (NULL);
1865 }
1866 
1867 static void
1868 xnf_tx_setup_offload(xnf_t *xnfp, xnf_txbuf_t *head,
1869     uint32_t cksum_flags, uint32_t lso_flags, uint32_t mss)
1870 {
1871 	if (lso_flags != 0) {
1872 		ASSERT3U(lso_flags, ==, HW_LSO);
1873 		ASSERT3P(head->tx_bdesc, ==, NULL);
1874 
1875 		head->tx_txreq.flags |= NETTXF_extra_info;
1876 		netif_extra_info_t *extra = &head->tx_extra;
1877 		extra->type = XEN_NETIF_EXTRA_TYPE_GSO;
1878 		extra->flags = 0;
1879 		extra->u.gso.type = XEN_NETIF_GSO_TYPE_TCPV4;
1880 		extra->u.gso.size = mss;
1881 		extra->u.gso.features = 0;
1882 		extra->u.gso.pad = 0;
1883 	} else if (cksum_flags != 0) {
1884 		ASSERT3U(cksum_flags, ==, HCK_FULLCKSUM);
1885 		/*
1886 		 * If the local protocol stack requests checksum
1887 		 * offload we set the 'checksum blank' flag,
1888 		 * indicating to the peer that we need the checksum
1889 		 * calculated for us.
1890 		 *
1891 		 * We _don't_ set the validated flag, because we haven't
1892 		 * validated that the data and the checksum match.
1893 		 *
1894 		 * Note: we already called xnf_pseudo_cksum() in
1895 		 * xnf_send(), so we just set the txreq flag here.
1896 		 */
1897 		head->tx_txreq.flags |= NETTXF_csum_blank;
1898 		xnfp->xnf_stat_tx_cksum_deferred++;
1899 	}
1900 }
1901 
1902 /*
1903  * Send packet mp. Called by the MAC framework.
1904  */
1905 static mblk_t *
1906 xnf_send(void *arg, mblk_t *mp)
1907 {
1908 	xnf_t *xnfp = arg;
1909 	xnf_txbuf_t *head;
1910 	mblk_t *ml;
1911 	int length;
1912 	int pages, chunks, slots, slots_free;
1913 	uint32_t cksum_flags, lso_flags, mss;
1914 	boolean_t pulledup = B_FALSE;
1915 	boolean_t force_copy = B_FALSE;
1916 
1917 	ASSERT3P(mp->b_next, ==, NULL);
1918 
1919 	mutex_enter(&xnfp->xnf_txlock);
1920 
1921 	/*
1922 	 * Wait until we are connected to the backend.
1923 	 */
1924 	while (!xnfp->xnf_connected)
1925 		cv_wait(&xnfp->xnf_cv_state, &xnfp->xnf_txlock);
1926 
1927 	/*
1928 	 * To simplify logic and be in sync with the rescheduling mechanism,
1929 	 * we require the maximum amount of slots that could be used by a
1930 	 * transaction to be free before proceeding. The only downside of doing
1931 	 * this is that it slightly reduces the effective size of the ring.
1932 	 */
1933 	slots_free = xnf_tx_slots_get(xnfp, XEN_MAX_SLOTS_PER_TX, B_FALSE);
1934 	if (slots_free < XEN_MAX_SLOTS_PER_TX) {
1935 		/*
1936 		 * We need to ask for a re-schedule later as the ring is full.
1937 		 */
1938 		mutex_enter(&xnfp->xnf_schedlock);
1939 		xnfp->xnf_need_sched = B_TRUE;
1940 		mutex_exit(&xnfp->xnf_schedlock);
1941 
1942 		xnfp->xnf_stat_tx_defer++;
1943 		mutex_exit(&xnfp->xnf_txlock);
1944 		return (mp);
1945 	}
1946 
1947 	/*
1948 	 * Get hw offload parameters.
1949 	 * This must be done before pulling up the mp as those parameters
1950 	 * are not copied over.
1951 	 */
1952 	mac_hcksum_get(mp, NULL, NULL, NULL, NULL, &cksum_flags);
1953 	mac_lso_get(mp, &mss, &lso_flags);
1954 
1955 	/*
1956 	 * XXX: fix MAC framework so that we can advertise support for
1957 	 * partial checksum for IPv4 only. This way we won't need to calculate
1958 	 * the pseudo header checksum ourselves.
1959 	 */
1960 	if (cksum_flags != 0) {
1961 		ASSERT3U(cksum_flags, ==, HCK_FULLCKSUM);
1962 		(void) xnf_pseudo_cksum(mp);
1963 	}
1964 
1965 pulledup:
1966 	for (ml = mp, pages = 0, chunks = 0, length = 0; ml != NULL;
1967 	    ml = ml->b_cont, chunks++) {
1968 		pages += xnf_mblk_pages(ml);
1969 		length += MBLKL(ml);
1970 	}
1971 	DTRACE_PROBE3(packet, int, length, int, chunks, int, pages);
1972 	DTRACE_PROBE3(lso, int, length, uint32_t, lso_flags, uint32_t, mss);
1973 
1974 	/*
1975 	 * If the ethernet header crosses a page boundary the packet
1976 	 * will be dropped by the backend. In practice it seems like
1977 	 * this happens fairly rarely so we'll do nothing unless the
1978 	 * packet is small enough to fit in a look-aside buffer.
1979 	 */
1980 	if (((uintptr_t)mp->b_rptr & PAGEOFFSET) +
1981 	    sizeof (struct ether_header) > PAGESIZE) {
1982 		xnfp->xnf_stat_tx_eth_hdr_split++;
1983 		if (length <= PAGESIZE)
1984 			force_copy = B_TRUE;
1985 	}
1986 
1987 	if (force_copy || (pages > 1 && !xnfp->xnf_be_tx_sg)) {
1988 		/*
1989 		 * If the packet spans several pages and scatter-gather is not
1990 		 * supported then use a look-aside buffer.
1991 		 */
1992 		ASSERT3U(length, <=, PAGESIZE);
1993 		head = xnf_mblk_copy(xnfp, mp);
1994 		if (head == NULL) {
1995 			dev_err(xnfp->xnf_devinfo, CE_WARN,
1996 			    "xnf_mblk_copy() failed");
1997 			goto drop;
1998 		}
1999 	} else {
2000 		/*
2001 		 * There's a limit for how many pages can be passed to the
2002 		 * backend. If we pass that limit, the packet will be dropped
2003 		 * and some backend implementations (e.g. Linux) could even
2004 		 * offline the interface.
2005 		 */
2006 		if (pages > XEN_MAX_TX_DATA_PAGES) {
2007 			if (pulledup) {
2008 				dev_err(xnfp->xnf_devinfo, CE_WARN,
2009 				    "too many pages, even after pullup: %d.",
2010 				    pages);
2011 				goto drop;
2012 			}
2013 
2014 			/*
2015 			 * Defragment packet if it spans too many pages.
2016 			 */
2017 			mblk_t *newmp = msgpullup(mp, -1);
2018 			freemsg(mp);
2019 			mp = newmp;
2020 			xnfp->xnf_stat_tx_pullup++;
2021 			pulledup = B_TRUE;
2022 			goto pulledup;
2023 		}
2024 
2025 		head = xnf_mblk_map(xnfp, mp, &slots);
2026 		if (head == NULL)
2027 			goto drop;
2028 
2029 		IMPLY(slots > 1, xnfp->xnf_be_tx_sg);
2030 	}
2031 
2032 	/*
2033 	 * Set tx_mp so that mblk is freed when the txbuf chain is freed.
2034 	 */
2035 	head->tx_mp = mp;
2036 
2037 	xnf_tx_setup_offload(xnfp, head, cksum_flags, lso_flags, mss);
2038 
2039 	/*
2040 	 * The first request must store the total length of the packet.
2041 	 */
2042 	head->tx_txreq.size = length;
2043 
2044 	/*
2045 	 * Push the packet we have prepared into the ring.
2046 	 */
2047 	xnf_tx_push_packet(xnfp, head);
2048 	xnfp->xnf_stat_opackets++;
2049 	xnfp->xnf_stat_obytes += length;
2050 
2051 	mutex_exit(&xnfp->xnf_txlock);
2052 	return (NULL);
2053 
2054 drop:
2055 	freemsg(mp);
2056 	xnfp->xnf_stat_tx_drop++;
2057 	mutex_exit(&xnfp->xnf_txlock);
2058 	return (NULL);
2059 }
2060 
2061 /*
2062  * Notification of RX packets. Currently no TX-complete interrupt is
2063  * used, as we clean the TX ring lazily.
2064  */
2065 static uint_t
2066 xnf_intr(caddr_t arg)
2067 {
2068 	xnf_t *xnfp = (xnf_t *)arg;
2069 	mblk_t *mp;
2070 	boolean_t need_sched, clean_ring;
2071 
2072 	mutex_enter(&xnfp->xnf_rxlock);
2073 
2074 	/*
2075 	 * Interrupts before we are connected are spurious.
2076 	 */
2077 	if (!xnfp->xnf_connected) {
2078 		mutex_exit(&xnfp->xnf_rxlock);
2079 		xnfp->xnf_stat_unclaimed_interrupts++;
2080 		return (DDI_INTR_UNCLAIMED);
2081 	}
2082 
2083 	/*
2084 	 * Receive side processing.
2085 	 */
2086 	do {
2087 		/*
2088 		 * Collect buffers from the ring.
2089 		 */
2090 		xnf_rx_collect(xnfp);
2091 
2092 		/*
2093 		 * Interrupt me when the next receive buffer is consumed.
2094 		 */
2095 		xnfp->xnf_rx_ring.sring->rsp_event =
2096 		    xnfp->xnf_rx_ring.rsp_cons + 1;
2097 		xen_mb();
2098 
2099 	} while (RING_HAS_UNCONSUMED_RESPONSES(&xnfp->xnf_rx_ring));
2100 
2101 	if (xnfp->xnf_rx_new_buffers_posted) {
2102 		boolean_t notify;
2103 
2104 		/*
2105 		 * Indicate to the peer that we have re-filled the
2106 		 * receive ring, if it cares.
2107 		 */
2108 		/* LINTED: constant in conditional context */
2109 		RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&xnfp->xnf_rx_ring, notify);
2110 		if (notify)
2111 			ec_notify_via_evtchn(xnfp->xnf_evtchn);
2112 		xnfp->xnf_rx_new_buffers_posted = B_FALSE;
2113 	}
2114 
2115 	mp = xnfp->xnf_rx_head;
2116 	xnfp->xnf_rx_head = xnfp->xnf_rx_tail = NULL;
2117 
2118 	xnfp->xnf_stat_interrupts++;
2119 	mutex_exit(&xnfp->xnf_rxlock);
2120 
2121 	if (mp != NULL)
2122 		mac_rx(xnfp->xnf_mh, NULL, mp);
2123 
2124 	/*
2125 	 * Transmit side processing.
2126 	 *
2127 	 * If a previous transmit attempt failed or we have pending
2128 	 * multicast requests, clean the ring.
2129 	 *
2130 	 * If we previously stalled transmission and cleaning produces
2131 	 * some free slots, tell upstream to attempt sending again.
2132 	 *
2133 	 * The odd style is to avoid acquiring xnf_txlock unless we
2134 	 * will actually look inside the tx machinery.
2135 	 */
2136 	mutex_enter(&xnfp->xnf_schedlock);
2137 	need_sched = xnfp->xnf_need_sched;
2138 	clean_ring = need_sched || (xnfp->xnf_pending_multicast > 0);
2139 	mutex_exit(&xnfp->xnf_schedlock);
2140 
2141 	if (clean_ring) {
2142 		int free_slots;
2143 
2144 		mutex_enter(&xnfp->xnf_txlock);
2145 		free_slots = xnf_tx_slots_get(xnfp, 0, B_FALSE);
2146 
2147 		if (need_sched && (free_slots >= XEN_MAX_SLOTS_PER_TX)) {
2148 			mutex_enter(&xnfp->xnf_schedlock);
2149 			xnfp->xnf_need_sched = B_FALSE;
2150 			mutex_exit(&xnfp->xnf_schedlock);
2151 
2152 			mac_tx_update(xnfp->xnf_mh);
2153 		}
2154 		mutex_exit(&xnfp->xnf_txlock);
2155 	}
2156 
2157 	return (DDI_INTR_CLAIMED);
2158 }
2159 
2160 /*
2161  *  xnf_start() -- start the board receiving and enable interrupts.
2162  */
2163 static int
2164 xnf_start(void *arg)
2165 {
2166 	xnf_t *xnfp = arg;
2167 
2168 #ifdef XNF_DEBUG
2169 	if (xnf_debug & XNF_DEBUG_TRACE)
2170 		printf("xnf%d start(0x%p)\n",
2171 		    ddi_get_instance(xnfp->xnf_devinfo), (void *)xnfp);
2172 #endif
2173 
2174 	mutex_enter(&xnfp->xnf_rxlock);
2175 	mutex_enter(&xnfp->xnf_txlock);
2176 
2177 	/* Accept packets from above. */
2178 	xnfp->xnf_running = B_TRUE;
2179 
2180 	mutex_exit(&xnfp->xnf_txlock);
2181 	mutex_exit(&xnfp->xnf_rxlock);
2182 
2183 	return (0);
2184 }
2185 
2186 /* xnf_stop() - disable hardware */
2187 static void
2188 xnf_stop(void *arg)
2189 {
2190 	xnf_t *xnfp = arg;
2191 
2192 #ifdef XNF_DEBUG
2193 	if (xnf_debug & XNF_DEBUG_TRACE)
2194 		printf("xnf%d stop(0x%p)\n",
2195 		    ddi_get_instance(xnfp->xnf_devinfo), (void *)xnfp);
2196 #endif
2197 
2198 	mutex_enter(&xnfp->xnf_rxlock);
2199 	mutex_enter(&xnfp->xnf_txlock);
2200 
2201 	xnfp->xnf_running = B_FALSE;
2202 
2203 	mutex_exit(&xnfp->xnf_txlock);
2204 	mutex_exit(&xnfp->xnf_rxlock);
2205 }
2206 
2207 /*
2208  * Hang buffer `bdesc' on the RX ring.
2209  */
2210 static void
2211 xnf_rxbuf_hang(xnf_t *xnfp, xnf_buf_t *bdesc)
2212 {
2213 	netif_rx_request_t *reqp;
2214 	RING_IDX hang_ix;
2215 
2216 	ASSERT(MUTEX_HELD(&xnfp->xnf_rxlock));
2217 
2218 	reqp = RING_GET_REQUEST(&xnfp->xnf_rx_ring,
2219 	    xnfp->xnf_rx_ring.req_prod_pvt);
2220 	hang_ix = (RING_IDX) (reqp - RING_GET_REQUEST(&xnfp->xnf_rx_ring, 0));
2221 	ASSERT(xnfp->xnf_rx_pkt_info[hang_ix] == NULL);
2222 
2223 	reqp->id = bdesc->id = hang_ix;
2224 	reqp->gref = bdesc->grant_ref;
2225 
2226 	xnfp->xnf_rx_pkt_info[hang_ix] = bdesc;
2227 	xnfp->xnf_rx_ring.req_prod_pvt++;
2228 
2229 	xnfp->xnf_rx_new_buffers_posted = B_TRUE;
2230 }
2231 
2232 /*
2233  * Receive an entire packet from the ring, starting from slot *consp.
2234  * prod indicates the slot of the latest response.
2235  * On return, *consp will point to the head of the next packet.
2236  *
2237  * Note: If slot prod was reached before we could gather a full packet, we will
2238  * drop the partial packet; this would most likely indicate a bug in either
2239  * the front-end or the back-end driver.
2240  *
2241  * An rx packet can consist of several fragments and thus span multiple slots.
2242  * Each fragment can contain up to 4k of data.
2243  *
2244  * A typical 9000 MTU packet with look like this:
2245  * +------+---------------------+-------------------+-----------------------+
2246  * | SLOT | TYPE                | CONTENTS          | FLAGS                 |
2247  * +------+---------------------+-------------------+-----------------------+
2248  * | 1    | netif_rx_response_t | 1st data fragment | more_data             |
2249  * +------+---------------------+-------------------+-----------------------+
2250  * | 2    | netif_rx_response_t | 2nd data fragment | more_data             |
2251  * +------+---------------------+-------------------+-----------------------+
2252  * | 3    | netif_rx_response_t | 3rd data fragment | [none]                |
2253  * +------+---------------------+-------------------+-----------------------+
2254  *
2255  * Fragments are chained by setting NETRXF_more_data in the previous
2256  * response's flags. If there are additional flags, such as
2257  * NETRXF_data_validated or NETRXF_extra_info, those should be set on the
2258  * first fragment.
2259  *
2260  * Sometimes extra info can be present. If so, it will follow the first
2261  * fragment, and NETRXF_extra_info flag will be set on the first response.
2262  * If LRO is set on a packet, it will be stored in the extra info. Conforming
2263  * to the spec, extra info can also be chained, but must all be present right
2264  * after the first fragment.
2265  *
2266  * Example of a packet with 2 extra infos:
2267  * +------+---------------------+-------------------+-----------------------+
2268  * | SLOT | TYPE                | CONTENTS          | FLAGS                 |
2269  * +------+---------------------+-------------------+-----------------------+
2270  * | 1    | netif_rx_response_t | 1st data fragment | extra_info, more_data |
2271  * +------+---------------------+-------------------+-----------------------+
2272  * | 2    | netif_extra_info_t  | 1st extra info    | EXTRA_FLAG_MORE       |
2273  * +------+---------------------+-------------------+-----------------------+
2274  * | 3    | netif_extra_info_t  | 2nd extra info    | [none]                |
2275  * +------+---------------------+-------------------+-----------------------+
2276  * | 4    | netif_rx_response_t | 2nd data fragment | more_data             |
2277  * +------+---------------------+-------------------+-----------------------+
2278  * | 5    | netif_rx_response_t | 3rd data fragment | more_data             |
2279  * +------+---------------------+-------------------+-----------------------+
2280  * | 6    | netif_rx_response_t | 4th data fragment | [none]                |
2281  * +------+---------------------+-------------------+-----------------------+
2282  *
2283  * In practice, the only extra we expect is for LRO, but only if we advertise
2284  * that we support it to the backend (xnf_enable_lro == TRUE).
2285  */
2286 static int
2287 xnf_rx_one_packet(xnf_t *xnfp, RING_IDX prod, RING_IDX *consp, mblk_t **mpp)
2288 {
2289 	mblk_t *head = NULL;
2290 	mblk_t *tail = NULL;
2291 	mblk_t *mp;
2292 	int error = 0;
2293 	RING_IDX cons = *consp;
2294 	netif_extra_info_t lro;
2295 	boolean_t is_lro = B_FALSE;
2296 	boolean_t is_extra = B_FALSE;
2297 
2298 	netif_rx_response_t rsp = *RING_GET_RESPONSE(&xnfp->xnf_rx_ring, cons);
2299 
2300 	boolean_t hwcsum = (rsp.flags & NETRXF_data_validated) != 0;
2301 	boolean_t more_data = (rsp.flags & NETRXF_more_data) != 0;
2302 	boolean_t more_extra = (rsp.flags & NETRXF_extra_info) != 0;
2303 
2304 	IMPLY(more_data, xnf_enable_rx_sg);
2305 
2306 	while (cons != prod) {
2307 		xnf_buf_t *bdesc;
2308 		int len, off;
2309 		int rxidx = cons & (NET_RX_RING_SIZE - 1);
2310 
2311 		bdesc = xnfp->xnf_rx_pkt_info[rxidx];
2312 		xnfp->xnf_rx_pkt_info[rxidx] = NULL;
2313 
2314 		if (is_extra) {
2315 			netif_extra_info_t *extra = (netif_extra_info_t *)&rsp;
2316 			/*
2317 			 * The only extra we expect is for LRO, and it should
2318 			 * only be present once.
2319 			 */
2320 			if (extra->type == XEN_NETIF_EXTRA_TYPE_GSO &&
2321 			    !is_lro) {
2322 				ASSERT(xnf_enable_lro);
2323 				lro = *extra;
2324 				is_lro = B_TRUE;
2325 				DTRACE_PROBE1(lro, netif_extra_info_t *, &lro);
2326 			} else {
2327 				dev_err(xnfp->xnf_devinfo, CE_WARN, "rx packet "
2328 				    "contains unexpected extra info of type %d",
2329 				    extra->type);
2330 				error = EINVAL;
2331 			}
2332 			more_extra =
2333 			    (extra->flags & XEN_NETIF_EXTRA_FLAG_MORE) != 0;
2334 
2335 			goto hang_buf;
2336 		}
2337 
2338 		ASSERT3U(bdesc->id, ==, rsp.id);
2339 
2340 		/*
2341 		 * status stores packet length when >= 0, or errors when < 0.
2342 		 */
2343 		len = rsp.status;
2344 		off = rsp.offset;
2345 		more_data = (rsp.flags & NETRXF_more_data) != 0;
2346 
2347 		/*
2348 		 * sanity checks.
2349 		 */
2350 		if (!xnfp->xnf_running) {
2351 			error = EBUSY;
2352 		} else if (len <= 0) {
2353 			xnfp->xnf_stat_errrx++;
2354 
2355 			switch (len) {
2356 			case 0:
2357 				xnfp->xnf_stat_runt++;
2358 				break;
2359 			case NETIF_RSP_ERROR:
2360 				xnfp->xnf_stat_mac_rcv_error++;
2361 				break;
2362 			case NETIF_RSP_DROPPED:
2363 				xnfp->xnf_stat_norxbuf++;
2364 				break;
2365 			}
2366 			error = EINVAL;
2367 		} else if (bdesc->grant_ref == INVALID_GRANT_REF) {
2368 			dev_err(xnfp->xnf_devinfo, CE_WARN,
2369 			    "Bad rx grant reference, rsp id %d", rsp.id);
2370 			error = EINVAL;
2371 		} else if ((off + len) > PAGESIZE) {
2372 			dev_err(xnfp->xnf_devinfo, CE_WARN, "Rx packet crosses "
2373 			    "page boundary (offset %d, length %d)", off, len);
2374 			error = EINVAL;
2375 		}
2376 
2377 		if (error != 0) {
2378 			/*
2379 			 * If an error has been detected, we do not attempt
2380 			 * to read the data but we still need to replace
2381 			 * the rx bufs.
2382 			 */
2383 			goto hang_buf;
2384 		}
2385 
2386 		xnf_buf_t *nbuf = NULL;
2387 
2388 		/*
2389 		 * If the packet is below a pre-determined size we will
2390 		 * copy data out of the buf rather than replace it.
2391 		 */
2392 		if (len > xnf_rx_copy_limit)
2393 			nbuf = xnf_buf_get(xnfp, KM_NOSLEEP, B_FALSE);
2394 
2395 		if (nbuf != NULL) {
2396 			mp = desballoc((unsigned char *)bdesc->buf,
2397 			    bdesc->len, 0, &bdesc->free_rtn);
2398 
2399 			if (mp == NULL) {
2400 				xnfp->xnf_stat_rx_desballoc_fail++;
2401 				xnfp->xnf_stat_norxbuf++;
2402 				error = ENOMEM;
2403 				/*
2404 				 * we free the buf we just allocated as we
2405 				 * will re-hang the old buf.
2406 				 */
2407 				xnf_buf_put(xnfp, nbuf, B_FALSE);
2408 				goto hang_buf;
2409 			}
2410 
2411 			mp->b_rptr = mp->b_rptr + off;
2412 			mp->b_wptr = mp->b_rptr + len;
2413 
2414 			/*
2415 			 * Release the grant as the backend doesn't need to
2416 			 * access this buffer anymore and grants are scarce.
2417 			 */
2418 			(void) gnttab_end_foreign_access_ref(bdesc->grant_ref,
2419 			    0);
2420 			xnf_gref_put(xnfp, bdesc->grant_ref);
2421 			bdesc->grant_ref = INVALID_GRANT_REF;
2422 
2423 			bdesc = nbuf;
2424 		} else {
2425 			/*
2426 			 * We failed to allocate a new buf or decided to reuse
2427 			 * the old one. In either case we copy the data off it
2428 			 * and put it back into the ring.
2429 			 */
2430 			mp = allocb(len, 0);
2431 			if (mp == NULL) {
2432 				xnfp->xnf_stat_rx_allocb_fail++;
2433 				xnfp->xnf_stat_norxbuf++;
2434 				error = ENOMEM;
2435 				goto hang_buf;
2436 			}
2437 			bcopy(bdesc->buf + off, mp->b_wptr, len);
2438 			mp->b_wptr += len;
2439 		}
2440 
2441 		if (head == NULL)
2442 			head = mp;
2443 		else
2444 			tail->b_cont = mp;
2445 		tail = mp;
2446 
2447 hang_buf:
2448 		/*
2449 		 * No matter what happens, for each response we need to hang
2450 		 * a new buf on the rx ring. Put either the old one, or a new
2451 		 * one if the old one is borrowed by the kernel via desballoc().
2452 		 */
2453 		xnf_rxbuf_hang(xnfp, bdesc);
2454 		cons++;
2455 
2456 		/* next response is an extra */
2457 		is_extra = more_extra;
2458 
2459 		if (!more_data && !more_extra)
2460 			break;
2461 
2462 		/*
2463 		 * Note that since requests and responses are union'd on the
2464 		 * same ring, we copy the response to a local variable instead
2465 		 * of keeping a pointer. Otherwise xnf_rxbuf_hang() would have
2466 		 * overwritten contents of rsp.
2467 		 */
2468 		rsp = *RING_GET_RESPONSE(&xnfp->xnf_rx_ring, cons);
2469 	}
2470 
2471 	/*
2472 	 * Check that we do not get stuck in a loop.
2473 	 */
2474 	ASSERT3U(*consp, !=, cons);
2475 	*consp = cons;
2476 
2477 	/*
2478 	 * We ran out of responses but the flags indicate there is more data.
2479 	 */
2480 	if (more_data) {
2481 		dev_err(xnfp->xnf_devinfo, CE_WARN, "rx: need more fragments.");
2482 		error = EINVAL;
2483 	}
2484 	if (more_extra) {
2485 		dev_err(xnfp->xnf_devinfo, CE_WARN, "rx: need more fragments "
2486 		    "(extras).");
2487 		error = EINVAL;
2488 	}
2489 
2490 	/*
2491 	 * An error means the packet must be dropped. If we have already formed
2492 	 * a partial packet, then discard it.
2493 	 */
2494 	if (error != 0) {
2495 		if (head != NULL)
2496 			freemsg(head);
2497 		xnfp->xnf_stat_rx_drop++;
2498 		return (error);
2499 	}
2500 
2501 	ASSERT(head != NULL);
2502 
2503 	if (hwcsum) {
2504 		/*
2505 		 * If the peer says that the data has been validated then we
2506 		 * declare that the full checksum has been verified.
2507 		 *
2508 		 * We don't look at the "checksum blank" flag, and hence could
2509 		 * have a packet here that we are asserting is good with
2510 		 * a blank checksum.
2511 		 */
2512 		mac_hcksum_set(head, 0, 0, 0, 0, HCK_FULLCKSUM_OK);
2513 		xnfp->xnf_stat_rx_cksum_no_need++;
2514 	}
2515 
2516 	/* XXX: set lro info for packet once LRO is supported in OS. */
2517 
2518 	*mpp = head;
2519 
2520 	return (0);
2521 }
2522 
2523 /*
2524  * Collect packets from the RX ring, storing them in `xnfp' for later use.
2525  */
2526 static void
2527 xnf_rx_collect(xnf_t *xnfp)
2528 {
2529 	RING_IDX prod;
2530 
2531 	ASSERT(MUTEX_HELD(&xnfp->xnf_rxlock));
2532 
2533 	prod = xnfp->xnf_rx_ring.sring->rsp_prod;
2534 	/*
2535 	 * Ensure we see queued responses up to 'prod'.
2536 	 */
2537 	membar_consumer();
2538 
2539 	while (xnfp->xnf_rx_ring.rsp_cons != prod) {
2540 		mblk_t *mp;
2541 
2542 		/*
2543 		 * Collect a packet.
2544 		 * rsp_cons is updated inside xnf_rx_one_packet().
2545 		 */
2546 		int error = xnf_rx_one_packet(xnfp, prod,
2547 		    &xnfp->xnf_rx_ring.rsp_cons, &mp);
2548 		if (error == 0) {
2549 			xnfp->xnf_stat_ipackets++;
2550 			xnfp->xnf_stat_rbytes += xmsgsize(mp);
2551 
2552 			/*
2553 			 * Append the mblk to the rx list.
2554 			 */
2555 			if (xnfp->xnf_rx_head == NULL) {
2556 				ASSERT3P(xnfp->xnf_rx_tail, ==, NULL);
2557 				xnfp->xnf_rx_head = mp;
2558 			} else {
2559 				ASSERT(xnfp->xnf_rx_tail != NULL);
2560 				xnfp->xnf_rx_tail->b_next = mp;
2561 			}
2562 			xnfp->xnf_rx_tail = mp;
2563 		}
2564 	}
2565 }
2566 
2567 /*
2568  *  xnf_alloc_dma_resources() -- initialize the drivers structures
2569  */
2570 static int
2571 xnf_alloc_dma_resources(xnf_t *xnfp)
2572 {
2573 	dev_info_t 		*devinfo = xnfp->xnf_devinfo;
2574 	size_t			len;
2575 	ddi_dma_cookie_t	dma_cookie;
2576 	uint_t			ncookies;
2577 	int			rc;
2578 	caddr_t			rptr;
2579 
2580 	/*
2581 	 * The code below allocates all the DMA data structures that
2582 	 * need to be released when the driver is detached.
2583 	 *
2584 	 * Allocate page for the transmit descriptor ring.
2585 	 */
2586 	if (ddi_dma_alloc_handle(devinfo, &ringbuf_dma_attr,
2587 	    DDI_DMA_SLEEP, 0, &xnfp->xnf_tx_ring_dma_handle) != DDI_SUCCESS)
2588 		goto alloc_error;
2589 
2590 	if (ddi_dma_mem_alloc(xnfp->xnf_tx_ring_dma_handle,
2591 	    PAGESIZE, &accattr, DDI_DMA_CONSISTENT,
2592 	    DDI_DMA_SLEEP, 0, &rptr, &len,
2593 	    &xnfp->xnf_tx_ring_dma_acchandle) != DDI_SUCCESS) {
2594 		ddi_dma_free_handle(&xnfp->xnf_tx_ring_dma_handle);
2595 		xnfp->xnf_tx_ring_dma_handle = NULL;
2596 		goto alloc_error;
2597 	}
2598 
2599 	if ((rc = ddi_dma_addr_bind_handle(xnfp->xnf_tx_ring_dma_handle, NULL,
2600 	    rptr, PAGESIZE, DDI_DMA_RDWR | DDI_DMA_CONSISTENT,
2601 	    DDI_DMA_SLEEP, 0, &dma_cookie, &ncookies)) != DDI_DMA_MAPPED) {
2602 		ddi_dma_mem_free(&xnfp->xnf_tx_ring_dma_acchandle);
2603 		ddi_dma_free_handle(&xnfp->xnf_tx_ring_dma_handle);
2604 		xnfp->xnf_tx_ring_dma_handle = NULL;
2605 		xnfp->xnf_tx_ring_dma_acchandle = NULL;
2606 		if (rc == DDI_DMA_NORESOURCES)
2607 			goto alloc_error;
2608 		else
2609 			goto error;
2610 	}
2611 
2612 	ASSERT(ncookies == 1);
2613 	bzero(rptr, PAGESIZE);
2614 	/* LINTED: constant in conditional context */
2615 	SHARED_RING_INIT((netif_tx_sring_t *)rptr);
2616 	/* LINTED: constant in conditional context */
2617 	FRONT_RING_INIT(&xnfp->xnf_tx_ring, (netif_tx_sring_t *)rptr, PAGESIZE);
2618 	xnfp->xnf_tx_ring_phys_addr = dma_cookie.dmac_laddress;
2619 
2620 	/*
2621 	 * Allocate page for the receive descriptor ring.
2622 	 */
2623 	if (ddi_dma_alloc_handle(devinfo, &ringbuf_dma_attr,
2624 	    DDI_DMA_SLEEP, 0, &xnfp->xnf_rx_ring_dma_handle) != DDI_SUCCESS)
2625 		goto alloc_error;
2626 
2627 	if (ddi_dma_mem_alloc(xnfp->xnf_rx_ring_dma_handle,
2628 	    PAGESIZE, &accattr, DDI_DMA_CONSISTENT,
2629 	    DDI_DMA_SLEEP, 0, &rptr, &len,
2630 	    &xnfp->xnf_rx_ring_dma_acchandle) != DDI_SUCCESS) {
2631 		ddi_dma_free_handle(&xnfp->xnf_rx_ring_dma_handle);
2632 		xnfp->xnf_rx_ring_dma_handle = NULL;
2633 		goto alloc_error;
2634 	}
2635 
2636 	if ((rc = ddi_dma_addr_bind_handle(xnfp->xnf_rx_ring_dma_handle, NULL,
2637 	    rptr, PAGESIZE, DDI_DMA_RDWR | DDI_DMA_CONSISTENT,
2638 	    DDI_DMA_SLEEP, 0, &dma_cookie, &ncookies)) != DDI_DMA_MAPPED) {
2639 		ddi_dma_mem_free(&xnfp->xnf_rx_ring_dma_acchandle);
2640 		ddi_dma_free_handle(&xnfp->xnf_rx_ring_dma_handle);
2641 		xnfp->xnf_rx_ring_dma_handle = NULL;
2642 		xnfp->xnf_rx_ring_dma_acchandle = NULL;
2643 		if (rc == DDI_DMA_NORESOURCES)
2644 			goto alloc_error;
2645 		else
2646 			goto error;
2647 	}
2648 
2649 	ASSERT(ncookies == 1);
2650 	bzero(rptr, PAGESIZE);
2651 	/* LINTED: constant in conditional context */
2652 	SHARED_RING_INIT((netif_rx_sring_t *)rptr);
2653 	/* LINTED: constant in conditional context */
2654 	FRONT_RING_INIT(&xnfp->xnf_rx_ring, (netif_rx_sring_t *)rptr, PAGESIZE);
2655 	xnfp->xnf_rx_ring_phys_addr = dma_cookie.dmac_laddress;
2656 
2657 	return (DDI_SUCCESS);
2658 
2659 alloc_error:
2660 	cmn_err(CE_WARN, "xnf%d: could not allocate enough DMA memory",
2661 	    ddi_get_instance(xnfp->xnf_devinfo));
2662 error:
2663 	xnf_release_dma_resources(xnfp);
2664 	return (DDI_FAILURE);
2665 }
2666 
2667 /*
2668  * Release all DMA resources in the opposite order from acquisition
2669  */
2670 static void
2671 xnf_release_dma_resources(xnf_t *xnfp)
2672 {
2673 	int i;
2674 
2675 	/*
2676 	 * Free receive buffers which are currently associated with
2677 	 * descriptors.
2678 	 */
2679 	mutex_enter(&xnfp->xnf_rxlock);
2680 	for (i = 0; i < NET_RX_RING_SIZE; i++) {
2681 		xnf_buf_t *bp;
2682 
2683 		if ((bp = xnfp->xnf_rx_pkt_info[i]) == NULL)
2684 			continue;
2685 		xnfp->xnf_rx_pkt_info[i] = NULL;
2686 		xnf_buf_put(xnfp, bp, B_FALSE);
2687 	}
2688 	mutex_exit(&xnfp->xnf_rxlock);
2689 
2690 	/* Free the receive ring buffer. */
2691 	if (xnfp->xnf_rx_ring_dma_acchandle != NULL) {
2692 		(void) ddi_dma_unbind_handle(xnfp->xnf_rx_ring_dma_handle);
2693 		ddi_dma_mem_free(&xnfp->xnf_rx_ring_dma_acchandle);
2694 		ddi_dma_free_handle(&xnfp->xnf_rx_ring_dma_handle);
2695 		xnfp->xnf_rx_ring_dma_acchandle = NULL;
2696 	}
2697 	/* Free the transmit ring buffer. */
2698 	if (xnfp->xnf_tx_ring_dma_acchandle != NULL) {
2699 		(void) ddi_dma_unbind_handle(xnfp->xnf_tx_ring_dma_handle);
2700 		ddi_dma_mem_free(&xnfp->xnf_tx_ring_dma_acchandle);
2701 		ddi_dma_free_handle(&xnfp->xnf_tx_ring_dma_handle);
2702 		xnfp->xnf_tx_ring_dma_acchandle = NULL;
2703 	}
2704 
2705 }
2706 
2707 /*
2708  * Release any packets and associated structures used by the TX ring.
2709  */
2710 static void
2711 xnf_release_mblks(xnf_t *xnfp)
2712 {
2713 	RING_IDX i;
2714 	xnf_txid_t *tidp;
2715 
2716 	for (i = 0, tidp = &xnfp->xnf_tx_pkt_id[0];
2717 	    i < NET_TX_RING_SIZE;
2718 	    i++, tidp++) {
2719 		xnf_txbuf_t *txp = tidp->txbuf;
2720 
2721 		if (txp != NULL) {
2722 			ASSERT(txp->tx_mp != NULL);
2723 			freemsg(txp->tx_mp);
2724 
2725 			xnf_txid_put(xnfp, tidp);
2726 			kmem_cache_free(xnfp->xnf_tx_buf_cache, txp);
2727 		}
2728 	}
2729 }
2730 
2731 static int
2732 xnf_buf_constructor(void *buf, void *arg, int kmflag)
2733 {
2734 	int (*ddiflags)(caddr_t) = DDI_DMA_SLEEP;
2735 	xnf_buf_t *bdesc = buf;
2736 	xnf_t *xnfp = arg;
2737 	ddi_dma_cookie_t dma_cookie;
2738 	uint_t ncookies;
2739 	size_t len;
2740 
2741 	if (kmflag & KM_NOSLEEP)
2742 		ddiflags = DDI_DMA_DONTWAIT;
2743 
2744 	/* Allocate a DMA access handle for the buffer. */
2745 	if (ddi_dma_alloc_handle(xnfp->xnf_devinfo, &rx_buf_dma_attr,
2746 	    ddiflags, 0, &bdesc->dma_handle) != DDI_SUCCESS)
2747 		goto failure;
2748 
2749 	/* Allocate DMA-able memory for buffer. */
2750 	if (ddi_dma_mem_alloc(bdesc->dma_handle,
2751 	    PAGESIZE, &data_accattr, DDI_DMA_STREAMING, ddiflags, 0,
2752 	    &bdesc->buf, &len, &bdesc->acc_handle) != DDI_SUCCESS)
2753 		goto failure_1;
2754 
2755 	/* Bind to virtual address of buffer to get physical address. */
2756 	if (ddi_dma_addr_bind_handle(bdesc->dma_handle, NULL,
2757 	    bdesc->buf, len, DDI_DMA_RDWR | DDI_DMA_STREAMING,
2758 	    ddiflags, 0, &dma_cookie, &ncookies) != DDI_DMA_MAPPED)
2759 		goto failure_2;
2760 	ASSERT(ncookies == 1);
2761 
2762 	bdesc->free_rtn.free_func = xnf_buf_recycle;
2763 	bdesc->free_rtn.free_arg = (caddr_t)bdesc;
2764 	bdesc->xnfp = xnfp;
2765 	bdesc->buf_phys = dma_cookie.dmac_laddress;
2766 	bdesc->buf_mfn = pfn_to_mfn(xnf_btop(bdesc->buf_phys));
2767 	bdesc->len = dma_cookie.dmac_size;
2768 	bdesc->grant_ref = INVALID_GRANT_REF;
2769 	bdesc->gen = xnfp->xnf_gen;
2770 
2771 	atomic_inc_64(&xnfp->xnf_stat_buf_allocated);
2772 
2773 	return (0);
2774 
2775 failure_2:
2776 	ddi_dma_mem_free(&bdesc->acc_handle);
2777 
2778 failure_1:
2779 	ddi_dma_free_handle(&bdesc->dma_handle);
2780 
2781 failure:
2782 
2783 	ASSERT(kmflag & KM_NOSLEEP); /* Cannot fail for KM_SLEEP. */
2784 	return (-1);
2785 }
2786 
2787 static void
2788 xnf_buf_destructor(void *buf, void *arg)
2789 {
2790 	xnf_buf_t *bdesc = buf;
2791 	xnf_t *xnfp = arg;
2792 
2793 	(void) ddi_dma_unbind_handle(bdesc->dma_handle);
2794 	ddi_dma_mem_free(&bdesc->acc_handle);
2795 	ddi_dma_free_handle(&bdesc->dma_handle);
2796 
2797 	atomic_dec_64(&xnfp->xnf_stat_buf_allocated);
2798 }
2799 
2800 static xnf_buf_t *
2801 xnf_buf_get(xnf_t *xnfp, int flags, boolean_t readonly)
2802 {
2803 	grant_ref_t gref;
2804 	xnf_buf_t *bufp;
2805 
2806 	/*
2807 	 * Usually grant references are more scarce than memory, so we
2808 	 * attempt to acquire a grant reference first.
2809 	 */
2810 	gref = xnf_gref_get(xnfp);
2811 	if (gref == INVALID_GRANT_REF)
2812 		return (NULL);
2813 
2814 	bufp = kmem_cache_alloc(xnfp->xnf_buf_cache, flags);
2815 	if (bufp == NULL) {
2816 		xnf_gref_put(xnfp, gref);
2817 		return (NULL);
2818 	}
2819 
2820 	ASSERT3U(bufp->grant_ref, ==, INVALID_GRANT_REF);
2821 
2822 	bufp->grant_ref = gref;
2823 
2824 	if (bufp->gen != xnfp->xnf_gen)
2825 		xnf_buf_refresh(bufp);
2826 
2827 	gnttab_grant_foreign_access_ref(bufp->grant_ref,
2828 	    xvdi_get_oeid(bufp->xnfp->xnf_devinfo),
2829 	    bufp->buf_mfn, readonly ? 1 : 0);
2830 
2831 	atomic_inc_64(&xnfp->xnf_stat_buf_outstanding);
2832 
2833 	return (bufp);
2834 }
2835 
2836 static void
2837 xnf_buf_put(xnf_t *xnfp, xnf_buf_t *bufp, boolean_t readonly)
2838 {
2839 	if (bufp->grant_ref != INVALID_GRANT_REF) {
2840 		(void) gnttab_end_foreign_access_ref(
2841 		    bufp->grant_ref, readonly ? 1 : 0);
2842 		xnf_gref_put(xnfp, bufp->grant_ref);
2843 		bufp->grant_ref = INVALID_GRANT_REF;
2844 	}
2845 
2846 	kmem_cache_free(xnfp->xnf_buf_cache, bufp);
2847 
2848 	atomic_dec_64(&xnfp->xnf_stat_buf_outstanding);
2849 }
2850 
2851 /*
2852  * Refresh any cached data about a buffer after resume.
2853  */
2854 static void
2855 xnf_buf_refresh(xnf_buf_t *bdesc)
2856 {
2857 	bdesc->buf_mfn = pfn_to_mfn(xnf_btop(bdesc->buf_phys));
2858 	bdesc->gen = bdesc->xnfp->xnf_gen;
2859 }
2860 
2861 /*
2862  * Streams `freeb' routine for `xnf_buf_t' when used as transmit
2863  * look-aside buffers.
2864  */
2865 static void
2866 xnf_buf_recycle(xnf_buf_t *bdesc)
2867 {
2868 	xnf_t *xnfp = bdesc->xnfp;
2869 
2870 	xnf_buf_put(xnfp, bdesc, B_TRUE);
2871 }
2872 
2873 static int
2874 xnf_tx_buf_constructor(void *buf, void *arg, int kmflag)
2875 {
2876 	int (*ddiflags)(caddr_t) = DDI_DMA_SLEEP;
2877 	xnf_txbuf_t *txp = buf;
2878 	xnf_t *xnfp = arg;
2879 
2880 	if (kmflag & KM_NOSLEEP)
2881 		ddiflags = DDI_DMA_DONTWAIT;
2882 
2883 	if (ddi_dma_alloc_handle(xnfp->xnf_devinfo, &tx_buf_dma_attr,
2884 	    ddiflags, 0, &txp->tx_dma_handle) != DDI_SUCCESS) {
2885 		ASSERT(kmflag & KM_NOSLEEP); /* Cannot fail for KM_SLEEP. */
2886 		return (-1);
2887 	}
2888 
2889 	return (0);
2890 }
2891 
2892 static void
2893 xnf_tx_buf_destructor(void *buf, void *arg)
2894 {
2895 	_NOTE(ARGUNUSED(arg));
2896 	xnf_txbuf_t *txp = buf;
2897 
2898 	ddi_dma_free_handle(&txp->tx_dma_handle);
2899 }
2900 
2901 /*
2902  * Statistics.
2903  */
2904 static char *xnf_aux_statistics[] = {
2905 	"tx_cksum_deferred",
2906 	"rx_cksum_no_need",
2907 	"interrupts",
2908 	"unclaimed_interrupts",
2909 	"tx_pullup",
2910 	"tx_lookaside",
2911 	"tx_drop",
2912 	"tx_eth_hdr_split",
2913 	"buf_allocated",
2914 	"buf_outstanding",
2915 	"gref_outstanding",
2916 	"gref_failure",
2917 	"gref_peak",
2918 	"rx_allocb_fail",
2919 	"rx_desballoc_fail",
2920 };
2921 
2922 static int
2923 xnf_kstat_aux_update(kstat_t *ksp, int flag)
2924 {
2925 	xnf_t *xnfp;
2926 	kstat_named_t *knp;
2927 
2928 	if (flag != KSTAT_READ)
2929 		return (EACCES);
2930 
2931 	xnfp = ksp->ks_private;
2932 	knp = ksp->ks_data;
2933 
2934 	/*
2935 	 * Assignment order must match that of the names in
2936 	 * xnf_aux_statistics.
2937 	 */
2938 	(knp++)->value.ui64 = xnfp->xnf_stat_tx_cksum_deferred;
2939 	(knp++)->value.ui64 = xnfp->xnf_stat_rx_cksum_no_need;
2940 
2941 	(knp++)->value.ui64 = xnfp->xnf_stat_interrupts;
2942 	(knp++)->value.ui64 = xnfp->xnf_stat_unclaimed_interrupts;
2943 	(knp++)->value.ui64 = xnfp->xnf_stat_tx_pullup;
2944 	(knp++)->value.ui64 = xnfp->xnf_stat_tx_lookaside;
2945 	(knp++)->value.ui64 = xnfp->xnf_stat_tx_drop;
2946 	(knp++)->value.ui64 = xnfp->xnf_stat_tx_eth_hdr_split;
2947 
2948 	(knp++)->value.ui64 = xnfp->xnf_stat_buf_allocated;
2949 	(knp++)->value.ui64 = xnfp->xnf_stat_buf_outstanding;
2950 	(knp++)->value.ui64 = xnfp->xnf_stat_gref_outstanding;
2951 	(knp++)->value.ui64 = xnfp->xnf_stat_gref_failure;
2952 	(knp++)->value.ui64 = xnfp->xnf_stat_gref_peak;
2953 	(knp++)->value.ui64 = xnfp->xnf_stat_rx_allocb_fail;
2954 	(knp++)->value.ui64 = xnfp->xnf_stat_rx_desballoc_fail;
2955 
2956 	return (0);
2957 }
2958 
2959 static boolean_t
2960 xnf_kstat_init(xnf_t *xnfp)
2961 {
2962 	int nstat = sizeof (xnf_aux_statistics) /
2963 	    sizeof (xnf_aux_statistics[0]);
2964 	char **cp = xnf_aux_statistics;
2965 	kstat_named_t *knp;
2966 
2967 	/*
2968 	 * Create and initialise kstats.
2969 	 */
2970 	if ((xnfp->xnf_kstat_aux = kstat_create("xnf",
2971 	    ddi_get_instance(xnfp->xnf_devinfo),
2972 	    "aux_statistics", "net", KSTAT_TYPE_NAMED,
2973 	    nstat, 0)) == NULL)
2974 		return (B_FALSE);
2975 
2976 	xnfp->xnf_kstat_aux->ks_private = xnfp;
2977 	xnfp->xnf_kstat_aux->ks_update = xnf_kstat_aux_update;
2978 
2979 	knp = xnfp->xnf_kstat_aux->ks_data;
2980 	while (nstat > 0) {
2981 		kstat_named_init(knp, *cp, KSTAT_DATA_UINT64);
2982 
2983 		knp++;
2984 		cp++;
2985 		nstat--;
2986 	}
2987 
2988 	kstat_install(xnfp->xnf_kstat_aux);
2989 
2990 	return (B_TRUE);
2991 }
2992 
2993 static int
2994 xnf_stat(void *arg, uint_t stat, uint64_t *val)
2995 {
2996 	xnf_t *xnfp = arg;
2997 
2998 	mutex_enter(&xnfp->xnf_rxlock);
2999 	mutex_enter(&xnfp->xnf_txlock);
3000 
3001 #define	mac_stat(q, r)				\
3002 	case (MAC_STAT_##q):			\
3003 		*val = xnfp->xnf_stat_##r;	\
3004 		break
3005 
3006 #define	ether_stat(q, r)			\
3007 	case (ETHER_STAT_##q):			\
3008 		*val = xnfp->xnf_stat_##r;	\
3009 		break
3010 
3011 	switch (stat) {
3012 
3013 	mac_stat(IPACKETS, ipackets);
3014 	mac_stat(OPACKETS, opackets);
3015 	mac_stat(RBYTES, rbytes);
3016 	mac_stat(OBYTES, obytes);
3017 	mac_stat(NORCVBUF, norxbuf);
3018 	mac_stat(IERRORS, errrx);
3019 	mac_stat(NOXMTBUF, tx_defer);
3020 
3021 	ether_stat(MACRCV_ERRORS, mac_rcv_error);
3022 	ether_stat(TOOSHORT_ERRORS, runt);
3023 
3024 	/* always claim to be in full duplex mode */
3025 	case ETHER_STAT_LINK_DUPLEX:
3026 		*val = LINK_DUPLEX_FULL;
3027 		break;
3028 
3029 	/* always claim to be at 1Gb/s link speed */
3030 	case MAC_STAT_IFSPEED:
3031 		*val = 1000000000ull;
3032 		break;
3033 
3034 	default:
3035 		mutex_exit(&xnfp->xnf_txlock);
3036 		mutex_exit(&xnfp->xnf_rxlock);
3037 
3038 		return (ENOTSUP);
3039 	}
3040 
3041 #undef mac_stat
3042 #undef ether_stat
3043 
3044 	mutex_exit(&xnfp->xnf_txlock);
3045 	mutex_exit(&xnfp->xnf_rxlock);
3046 
3047 	return (0);
3048 }
3049 
3050 static int
3051 xnf_change_mtu(xnf_t *xnfp, uint32_t mtu)
3052 {
3053 	if (mtu > ETHERMTU) {
3054 		if (!xnf_enable_tx_sg) {
3055 			dev_err(xnfp->xnf_devinfo, CE_WARN, "MTU limited to %d "
3056 			    "because scatter-gather is disabled for transmit "
3057 			    "in driver settings", ETHERMTU);
3058 			return (EINVAL);
3059 		} else if (!xnf_enable_rx_sg) {
3060 			dev_err(xnfp->xnf_devinfo, CE_WARN, "MTU limited to %d "
3061 			    "because scatter-gather is disabled for receive "
3062 			    "in driver settings", ETHERMTU);
3063 			return (EINVAL);
3064 		} else if (!xnfp->xnf_be_tx_sg) {
3065 			dev_err(xnfp->xnf_devinfo, CE_WARN, "MTU limited to %d "
3066 			    "because backend doesn't support scatter-gather",
3067 			    ETHERMTU);
3068 			return (EINVAL);
3069 		}
3070 		if (mtu > XNF_MAXPKT)
3071 			return (EINVAL);
3072 	}
3073 	int error = mac_maxsdu_update(xnfp->xnf_mh, mtu);
3074 	if (error == 0)
3075 		xnfp->xnf_mtu = mtu;
3076 
3077 	return (error);
3078 }
3079 
3080 /*ARGSUSED*/
3081 static int
3082 xnf_getprop(void *data, const char *prop_name, mac_prop_id_t prop_id,
3083     uint_t prop_val_size, void *prop_val)
3084 {
3085 	xnf_t *xnfp = data;
3086 
3087 	switch (prop_id) {
3088 	case MAC_PROP_MTU:
3089 		ASSERT(prop_val_size >= sizeof (uint32_t));
3090 		bcopy(&xnfp->xnf_mtu, prop_val, sizeof (uint32_t));
3091 		break;
3092 	default:
3093 		return (ENOTSUP);
3094 	}
3095 	return (0);
3096 }
3097 
3098 /*ARGSUSED*/
3099 static int
3100 xnf_setprop(void *data, const char *prop_name, mac_prop_id_t prop_id,
3101     uint_t prop_val_size, const void *prop_val)
3102 {
3103 	xnf_t *xnfp = data;
3104 	uint32_t new_mtu;
3105 	int error;
3106 
3107 	switch (prop_id) {
3108 	case MAC_PROP_MTU:
3109 		ASSERT(prop_val_size >= sizeof (uint32_t));
3110 		bcopy(prop_val, &new_mtu, sizeof (new_mtu));
3111 		error = xnf_change_mtu(xnfp, new_mtu);
3112 		break;
3113 	default:
3114 		return (ENOTSUP);
3115 	}
3116 
3117 	return (error);
3118 }
3119 
3120 /*ARGSUSED*/
3121 static void
3122 xnf_propinfo(void *data, const char *prop_name, mac_prop_id_t prop_id,
3123     mac_prop_info_handle_t prop_handle)
3124 {
3125 	switch (prop_id) {
3126 	case MAC_PROP_MTU:
3127 		mac_prop_info_set_range_uint32(prop_handle, 0, XNF_MAXPKT);
3128 		break;
3129 	default:
3130 		break;
3131 	}
3132 }
3133 
3134 static boolean_t
3135 xnf_getcapab(void *arg, mac_capab_t cap, void *cap_data)
3136 {
3137 	xnf_t *xnfp = arg;
3138 
3139 	switch (cap) {
3140 	case MAC_CAPAB_HCKSUM: {
3141 		uint32_t *capab = cap_data;
3142 
3143 		/*
3144 		 * Whilst the flag used to communicate with the IO
3145 		 * domain is called "NETTXF_csum_blank", the checksum
3146 		 * in the packet must contain the pseudo-header
3147 		 * checksum and not zero.
3148 		 *
3149 		 * To help out the IO domain, we might use
3150 		 * HCKSUM_INET_PARTIAL. Unfortunately our stack will
3151 		 * then use checksum offload for IPv6 packets, which
3152 		 * the IO domain can't handle.
3153 		 *
3154 		 * As a result, we declare outselves capable of
3155 		 * HCKSUM_INET_FULL_V4. This means that we receive
3156 		 * IPv4 packets from the stack with a blank checksum
3157 		 * field and must insert the pseudo-header checksum
3158 		 * before passing the packet to the IO domain.
3159 		 */
3160 		*capab = HCKSUM_INET_FULL_V4;
3161 
3162 		/*
3163 		 * TODO: query the "feature-ipv6-csum-offload" capability.
3164 		 * If enabled, that could allow us to use HCKSUM_INET_PARTIAL.
3165 		 */
3166 
3167 		break;
3168 	}
3169 	case MAC_CAPAB_LSO: {
3170 		if (!xnfp->xnf_be_lso)
3171 			return (B_FALSE);
3172 
3173 		mac_capab_lso_t *lso = cap_data;
3174 		lso->lso_flags = LSO_TX_BASIC_TCP_IPV4;
3175 		lso->lso_basic_tcp_ipv4.lso_max = IP_MAXPACKET;
3176 		break;
3177 	}
3178 	default:
3179 		return (B_FALSE);
3180 	}
3181 
3182 	return (B_TRUE);
3183 }
3184 
3185 /*
3186  * The state of the peer has changed - react accordingly.
3187  */
3188 static void
3189 oe_state_change(dev_info_t *dip, ddi_eventcookie_t id,
3190     void *arg, void *impl_data)
3191 {
3192 	_NOTE(ARGUNUSED(id, arg));
3193 	xnf_t *xnfp = ddi_get_driver_private(dip);
3194 	XenbusState new_state = *(XenbusState *)impl_data;
3195 
3196 	ASSERT(xnfp != NULL);
3197 
3198 	switch (new_state) {
3199 	case XenbusStateUnknown:
3200 	case XenbusStateInitialising:
3201 	case XenbusStateInitialised:
3202 	case XenbusStateClosing:
3203 	case XenbusStateClosed:
3204 	case XenbusStateReconfiguring:
3205 	case XenbusStateReconfigured:
3206 		break;
3207 
3208 	case XenbusStateInitWait:
3209 		xnf_read_config(xnfp);
3210 
3211 		if (!xnfp->xnf_be_rx_copy) {
3212 			cmn_err(CE_WARN,
3213 			    "The xnf driver requires a dom0 that "
3214 			    "supports 'feature-rx-copy'.");
3215 			(void) xvdi_switch_state(xnfp->xnf_devinfo,
3216 			    XBT_NULL, XenbusStateClosed);
3217 			break;
3218 		}
3219 
3220 		/*
3221 		 * Connect to the backend.
3222 		 */
3223 		xnf_be_connect(xnfp);
3224 
3225 		/*
3226 		 * Our MAC address as discovered by xnf_read_config().
3227 		 */
3228 		mac_unicst_update(xnfp->xnf_mh, xnfp->xnf_mac_addr);
3229 
3230 		/*
3231 		 * We do not know if some features such as LSO are supported
3232 		 * until we connect to the backend. We request the MAC layer
3233 		 * to poll our capabilities again.
3234 		 */
3235 		mac_capab_update(xnfp->xnf_mh);
3236 
3237 		break;
3238 
3239 	case XenbusStateConnected:
3240 		mutex_enter(&xnfp->xnf_rxlock);
3241 		mutex_enter(&xnfp->xnf_txlock);
3242 
3243 		xnfp->xnf_connected = B_TRUE;
3244 		/*
3245 		 * Wake up any threads waiting to send data to
3246 		 * backend.
3247 		 */
3248 		cv_broadcast(&xnfp->xnf_cv_state);
3249 
3250 		mutex_exit(&xnfp->xnf_txlock);
3251 		mutex_exit(&xnfp->xnf_rxlock);
3252 
3253 		/*
3254 		 * Kick the peer in case it missed any transmits
3255 		 * request in the TX ring.
3256 		 */
3257 		ec_notify_via_evtchn(xnfp->xnf_evtchn);
3258 
3259 		/*
3260 		 * There may already be completed receive requests in
3261 		 * the ring sent by backend after it gets connected
3262 		 * but before we see its state change here, so we call
3263 		 * xnf_intr() to handle them, if any.
3264 		 */
3265 		(void) xnf_intr((caddr_t)xnfp);
3266 
3267 		/*
3268 		 * Mark the link up now that we are connected.
3269 		 */
3270 		mac_link_update(xnfp->xnf_mh, LINK_STATE_UP);
3271 
3272 		/*
3273 		 * Tell the backend about the multicast addresses in
3274 		 * which we are interested.
3275 		 */
3276 		mac_multicast_refresh(xnfp->xnf_mh, NULL, xnfp, B_TRUE);
3277 
3278 		break;
3279 
3280 	default:
3281 		break;
3282 	}
3283 }
3284