xref: /illumos-gate/usr/src/uts/intel/io/viona/viona_tx.c (revision 06ca4e396ecc93ed45a333664a51558b7e84e8f5)
1 /*
2  * Copyright (c) 2013  Chris Torek <torek @ torek net>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 /*
27  * This file and its contents are supplied under the terms of the
28  * Common Development and Distribution License ("CDDL"), version 1.0.
29  * You may only use this file in accordance with the terms of version
30  * 1.0 of the CDDL.
31  *
32  * A full copy of the text of the CDDL should have accompanied this
33  * source.  A copy of the CDDL is also available via the Internet at
34  * http://www.illumos.org/license/CDDL.
35  *
36  * Copyright 2015 Pluribus Networks Inc.
37  * Copyright 2019 Joyent, Inc.
38  * Copyright 2022 Oxide Computer Company
39  */
40 
41 
42 #include <sys/types.h>
43 #include <sys/smt.h>
44 #include <sys/strsubr.h>
45 
46 #include <sys/pattr.h>
47 #include <sys/dlpi.h>
48 #include <inet/ip.h>
49 #include <inet/ip_impl.h>
50 
51 #include "viona_impl.h"
52 
53 #define	BNXE_NIC_DRIVER		"bnxe"
54 
55 /*
56  * copy tx mbufs from virtio ring to avoid necessitating a wait for packet
57  * transmission to free resources.
58  */
59 kmutex_t viona_force_copy_lock;
60 static enum viona_force_copy {
61 	VFC_UNINITALIZED	= 0,
62 	VFC_COPY_UNEEDED	= 1,
63 	VFC_COPY_REQUIRED	= 2,
64 } viona_force_copy_state = VFC_UNINITALIZED;
65 
66 struct viona_desb {
67 	frtn_t			d_frtn;
68 	viona_vring_t		*d_ring;
69 	uint_t			d_ref;
70 	uint32_t		d_len;
71 	uint16_t		d_cookie;
72 	uchar_t			*d_headers;
73 	vmm_page_t		*d_pages;
74 };
75 
76 static void viona_tx(viona_link_t *, viona_vring_t *);
77 static void viona_desb_release(viona_desb_t *);
78 
79 
80 static void
81 viona_tx_wait_outstanding(viona_vring_t *ring)
82 {
83 	ASSERT(MUTEX_HELD(&ring->vr_lock));
84 
85 	while (ring->vr_xfer_outstanding != 0) {
86 		/*
87 		 * Paying heed to signals is counterproductive here.  This is a
88 		 * very tight loop if pending transfers take an extended amount
89 		 * of time to be reclaimed while the host process is exiting.
90 		 */
91 		cv_wait(&ring->vr_cv, &ring->vr_lock);
92 	}
93 }
94 
95 /*
96  * Check if full TX packet copying is needed.  This should not be called from
97  * viona attach()/detach() context.
98  */
99 static boolean_t
100 viona_tx_copy_needed(void)
101 {
102 	boolean_t result;
103 
104 	mutex_enter(&viona_force_copy_lock);
105 	if (viona_force_copy_state == VFC_UNINITALIZED) {
106 		major_t bnxe_major;
107 
108 		/*
109 		 * The original code for viona featured an explicit check for
110 		 * the bnxe driver which, when found present, necessitated that
111 		 * all transmissions be copied into their own mblks instead of
112 		 * passing guest memory to the underlying device.
113 		 *
114 		 * The motivations for this are unclear, but until it can be
115 		 * proven unnecessary, the check lives on.
116 		 */
117 		viona_force_copy_state = VFC_COPY_UNEEDED;
118 		if ((bnxe_major = ddi_name_to_major(BNXE_NIC_DRIVER))
119 		    != DDI_MAJOR_T_NONE) {
120 			if (ddi_hold_installed_driver(bnxe_major) != NULL) {
121 				viona_force_copy_state = VFC_COPY_REQUIRED;
122 				ddi_rele_driver(bnxe_major);
123 			}
124 		}
125 	}
126 	result = (viona_force_copy_state == VFC_COPY_REQUIRED);
127 	mutex_exit(&viona_force_copy_lock);
128 
129 	return (result);
130 }
131 
132 void
133 viona_tx_ring_alloc(viona_vring_t *ring, const uint16_t qsz)
134 {
135 	/* Allocate desb handles for TX ring if packet copying not disabled */
136 	if (!viona_tx_copy_needed()) {
137 		viona_desb_t *dp;
138 
139 		dp = kmem_zalloc(sizeof (viona_desb_t) * qsz, KM_SLEEP);
140 		ring->vr_txdesb = dp;
141 		for (uint_t i = 0; i < qsz; i++, dp++) {
142 			dp->d_frtn.free_func = viona_desb_release;
143 			dp->d_frtn.free_arg = (void *)dp;
144 			dp->d_ring = ring;
145 			dp->d_headers = kmem_zalloc(VIONA_MAX_HDRS_LEN,
146 			    KM_SLEEP);
147 		}
148 	}
149 
150 	/* Allocate ring-sized iovec buffers for TX */
151 	ring->vr_txiov = kmem_alloc(sizeof (struct iovec) * qsz, KM_SLEEP);
152 }
153 
154 void
155 viona_tx_ring_free(viona_vring_t *ring, const uint16_t qsz)
156 {
157 	if (ring->vr_txdesb != NULL) {
158 		viona_desb_t *dp = ring->vr_txdesb;
159 
160 		for (uint_t i = 0; i < qsz; i++, dp++) {
161 			kmem_free(dp->d_headers, VIONA_MAX_HDRS_LEN);
162 		}
163 		kmem_free(ring->vr_txdesb, sizeof (viona_desb_t) * qsz);
164 		ring->vr_txdesb = NULL;
165 	}
166 
167 	if (ring->vr_txiov != NULL) {
168 		kmem_free(ring->vr_txiov, sizeof (struct iovec) * qsz);
169 		ring->vr_txiov = NULL;
170 	}
171 }
172 
173 static void
174 viona_tx_done(viona_vring_t *ring, uint32_t len, uint16_t cookie)
175 {
176 	vq_pushchain(ring, len, cookie);
177 
178 	membar_enter();
179 	viona_intr_ring(ring, B_FALSE);
180 }
181 
182 #define	TX_BURST_THRESH	32
183 
184 void
185 viona_worker_tx(viona_vring_t *ring, viona_link_t *link)
186 {
187 	(void) thread_vsetname(curthread, "viona_tx_%p", ring);
188 
189 	ASSERT(MUTEX_HELD(&ring->vr_lock));
190 	ASSERT3U(ring->vr_state, ==, VRS_RUN);
191 
192 	mutex_exit(&ring->vr_lock);
193 
194 	for (;;) {
195 		uint_t ntx = 0, burst = 0;
196 
197 		viona_ring_disable_notify(ring);
198 		while (viona_ring_num_avail(ring) != 0) {
199 			viona_tx(link, ring);
200 			ntx++;
201 			burst++;
202 
203 			/*
204 			 * It is advantageous for throughput to keep this
205 			 * transmission loop tight, but periodic breaks to
206 			 * check for other events are of value too.
207 			 */
208 			if (burst >= TX_BURST_THRESH) {
209 				mutex_enter(&ring->vr_lock);
210 				const bool need_bail = vring_need_bail(ring);
211 				mutex_exit(&ring->vr_lock);
212 
213 				if (need_bail) {
214 					break;
215 				}
216 				burst = 0;
217 			}
218 		}
219 
220 		VIONA_PROBE2(tx, viona_link_t *, link, uint_t, ntx);
221 
222 		/*
223 		 * Check for available descriptors on the ring once more in
224 		 * case a late addition raced with the NO_NOTIFY flag toggle.
225 		 *
226 		 * The barrier ensures that visibility of the no-notify
227 		 * store does not cross the viona_ring_num_avail() check below.
228 		 */
229 		viona_ring_enable_notify(ring);
230 		membar_enter();
231 
232 		if (viona_ring_num_avail(ring) == 0 &&
233 		    (link->l_features & VIRTIO_F_RING_NOTIFY_ON_EMPTY) != 0) {
234 			/*
235 			 * The NOTIFY_ON_EMPTY interrupt should not pay heed to
236 			 * the presence of AVAIL_NO_INTERRUPT.
237 			 */
238 			viona_intr_ring(ring, B_TRUE);
239 		}
240 
241 		mutex_enter(&ring->vr_lock);
242 		for (;;) {
243 			if (vring_need_bail(ring)) {
244 				ring->vr_state = VRS_STOP;
245 				viona_tx_wait_outstanding(ring);
246 				return;
247 			}
248 
249 			if (vmm_drv_lease_expired(ring->vr_lease)) {
250 				ring->vr_state_flags |= VRSF_RENEW;
251 				/*
252 				 * When renewing the lease for the ring, no TX
253 				 * frames may be outstanding, as they contain
254 				 * references to guest memory.
255 				 */
256 				viona_tx_wait_outstanding(ring);
257 
258 				const boolean_t renewed =
259 				    viona_ring_lease_renew(ring);
260 				ring->vr_state_flags &= ~VRSF_RENEW;
261 
262 				if (!renewed) {
263 					/* stop ring on failed renewal */
264 					ring->vr_state = VRS_STOP;
265 					return;
266 				}
267 			}
268 
269 			if (viona_ring_num_avail(ring) != 0) {
270 				break;
271 			}
272 
273 			/* Wait for further activity on the ring */
274 			(void) cv_wait_sig(&ring->vr_cv, &ring->vr_lock);
275 		}
276 		mutex_exit(&ring->vr_lock);
277 	}
278 	/* UNREACHABLE */
279 }
280 
281 static void
282 viona_desb_release(viona_desb_t *dp)
283 {
284 	viona_vring_t *ring = dp->d_ring;
285 	uint_t ref;
286 	uint32_t len;
287 	uint16_t cookie;
288 
289 	ref = atomic_dec_uint_nv(&dp->d_ref);
290 	if (ref > 1) {
291 		return;
292 	}
293 
294 	/*
295 	 * The desb corresponding to this index must be ready for reuse before
296 	 * the descriptor is returned to the guest via the 'used' ring.
297 	 */
298 	len = dp->d_len;
299 	cookie = dp->d_cookie;
300 	dp->d_len = 0;
301 	dp->d_cookie = 0;
302 	vmm_drv_page_release_chain(dp->d_pages);
303 	dp->d_pages = NULL;
304 
305 	/*
306 	 * Ensure all other changes to the desb are visible prior to zeroing its
307 	 * refcount, signifying its readiness for reuse.
308 	 */
309 	membar_exit();
310 	dp->d_ref = 0;
311 
312 	viona_tx_done(ring, len, cookie);
313 
314 	mutex_enter(&ring->vr_lock);
315 	if ((--ring->vr_xfer_outstanding) == 0) {
316 		cv_broadcast(&ring->vr_cv);
317 	}
318 	mutex_exit(&ring->vr_lock);
319 }
320 
321 static boolean_t
322 viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
323     mblk_t *mp, uint32_t len)
324 {
325 	viona_link_t *link = ring->vr_link;
326 	const struct ether_header *eth;
327 	uint_t eth_len = sizeof (struct ether_header);
328 	ushort_t ftype;
329 	ipha_t *ipha = NULL;
330 	uint8_t ipproto = IPPROTO_NONE; /* NONE is not exactly right, but ok */
331 	uint16_t flags = 0;
332 	const uint_t csum_start = hdr->vrh_csum_start;
333 	const uint_t csum_stuff = hdr->vrh_csum_offset + csum_start;
334 
335 	/*
336 	 * Validate that the checksum offsets provided by the guest are within
337 	 * the bounds of the packet.  Additionally, ensure that the checksum
338 	 * contents field is within the headers mblk copied by viona_tx().
339 	 */
340 	if (csum_start >= len || csum_start < eth_len || csum_stuff >= len ||
341 	    (csum_stuff + sizeof (uint16_t)) > MBLKL(mp)) {
342 		VIONA_PROBE2(fail_hcksum, viona_link_t *, link, mblk_t *, mp);
343 		VIONA_RING_STAT_INCR(ring, fail_hcksum);
344 		return (B_FALSE);
345 	}
346 
347 	/*
348 	 * This is guaranteed to be safe thanks to the header copying
349 	 * done in viona_tx().
350 	 */
351 	eth = (const struct ether_header *)mp->b_rptr;
352 	ftype = ntohs(eth->ether_type);
353 
354 	if (ftype == ETHERTYPE_VLAN) {
355 		const struct ether_vlan_header *veth;
356 
357 		/* punt on QinQ for now */
358 		eth_len = sizeof (struct ether_vlan_header);
359 		veth = (const struct ether_vlan_header *)eth;
360 		ftype = ntohs(veth->ether_type);
361 	}
362 
363 	if (ftype == ETHERTYPE_IP) {
364 		ipha = (ipha_t *)(mp->b_rptr + eth_len);
365 
366 		ipproto = ipha->ipha_protocol;
367 	} else if (ftype == ETHERTYPE_IPV6) {
368 		ip6_t *ip6h = (ip6_t *)(mp->b_rptr + eth_len);
369 
370 		ipproto = ip6h->ip6_nxt;
371 	}
372 
373 	/*
374 	 * We ignore hdr_len because the spec says it can't be
375 	 * trusted. Besides, our own stack will determine the header
376 	 * boundary.
377 	 */
378 	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0 &&
379 	    (hdr->vrh_gso_type & VIRTIO_NET_HDR_GSO_TCPV4) != 0 &&
380 	    ftype == ETHERTYPE_IP) {
381 		uint16_t	*cksump;
382 		uint32_t	cksum;
383 		ipaddr_t	src = ipha->ipha_src;
384 		ipaddr_t	dst = ipha->ipha_dst;
385 
386 		/*
387 		 * Our native IP stack doesn't set the L4 length field
388 		 * of the pseudo header when LSO is in play. Other IP
389 		 * stacks, e.g. Linux, do include the length field.
390 		 * This is a problem because the hardware expects that
391 		 * the length field is not set. When it is set it will
392 		 * cause an incorrect TCP checksum to be generated.
393 		 * The reason this works in Linux is because Linux
394 		 * corrects the pseudo-header checksum in the driver
395 		 * code. In order to get the correct HW checksum we
396 		 * need to assume the guest's IP stack gave us a bogus
397 		 * TCP partial checksum and calculate it ourselves.
398 		 */
399 		cksump = IPH_TCPH_CHECKSUMP(ipha, IPH_HDR_LENGTH(ipha));
400 		cksum = IP_TCP_CSUM_COMP;
401 		cksum += (dst >> 16) + (dst & 0xFFFF) +
402 		    (src >> 16) + (src & 0xFFFF);
403 		cksum = (cksum & 0xFFFF) + (cksum >> 16);
404 		*(cksump) = (cksum & 0xFFFF) + (cksum >> 16);
405 
406 		/*
407 		 * Since viona is a "legacy device", the data stored
408 		 * by the driver will be in the guest's native endian
409 		 * format (see sections 2.4.3 and 5.1.6.1 of the
410 		 * VIRTIO 1.0 spec for more info). At this time the
411 		 * only guests using viona are x86 and we can assume
412 		 * little-endian.
413 		 */
414 		lso_info_set(mp, LE_16(hdr->vrh_gso_size), HW_LSO);
415 
416 		/*
417 		 * Hardware, like ixgbe, expects the client to request
418 		 * IP header checksum offload if it's sending LSO (see
419 		 * ixgbe_get_context()). Unfortunately, virtio makes
420 		 * no allowances for negotiating IP header checksum
421 		 * and HW offload, only TCP checksum. We add the flag
422 		 * and zero-out the checksum field. This mirrors the
423 		 * behavior of our native IP stack (which does this in
424 		 * the interest of HW that expects the field to be
425 		 * zero).
426 		 */
427 		flags |= HCK_IPV4_HDRCKSUM;
428 		ipha->ipha_hdr_checksum = 0;
429 	}
430 
431 	/*
432 	 * Use DB_CKSUMFLAGS instead of mac_hcksum_get() to make sure
433 	 * HW_LSO, if present, is not lost.
434 	 */
435 	flags |= DB_CKSUMFLAGS(mp);
436 
437 	/*
438 	 * Partial checksum support from the NIC is ideal, since it most
439 	 * closely maps to the interface defined by virtio.
440 	 */
441 	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0 &&
442 	    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
443 		/*
444 		 * MAC expects these offsets to be relative to the
445 		 * start of the L3 header rather than the L2 frame.
446 		 */
447 		flags |= HCK_PARTIALCKSUM;
448 		mac_hcksum_set(mp, csum_start - eth_len, csum_stuff - eth_len,
449 		    len - eth_len, 0, flags);
450 		return (B_TRUE);
451 	}
452 
453 	/*
454 	 * Without partial checksum support, look to the L3/L4 protocol
455 	 * information to see if the NIC can handle it.  If not, the
456 	 * checksum will need to calculated inline.
457 	 */
458 	if (ftype == ETHERTYPE_IP) {
459 		if ((link->l_cap_csum & HCKSUM_INET_FULL_V4) != 0 &&
460 		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
461 			uint16_t *csump = (uint16_t *)(mp->b_rptr + csum_stuff);
462 			*csump = 0;
463 			flags |= HCK_FULLCKSUM;
464 			mac_hcksum_set(mp, 0, 0, 0, 0, flags);
465 			return (B_TRUE);
466 		}
467 
468 		/* XXX: Implement manual fallback checksumming? */
469 		VIONA_PROBE2(fail_hcksum, viona_link_t *, link, mblk_t *, mp);
470 		VIONA_RING_STAT_INCR(ring, fail_hcksum);
471 		return (B_FALSE);
472 	} else if (ftype == ETHERTYPE_IPV6) {
473 		if ((link->l_cap_csum & HCKSUM_INET_FULL_V6) != 0 &&
474 		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
475 			uint16_t *csump = (uint16_t *)(mp->b_rptr + csum_stuff);
476 			*csump = 0;
477 			flags |= HCK_FULLCKSUM;
478 			mac_hcksum_set(mp, 0, 0, 0, 0, flags);
479 			return (B_TRUE);
480 		}
481 
482 		/* XXX: Implement manual fallback checksumming? */
483 		VIONA_PROBE2(fail_hcksum6, viona_link_t *, link, mblk_t *, mp);
484 		VIONA_RING_STAT_INCR(ring, fail_hcksum6);
485 		return (B_FALSE);
486 	}
487 
488 	/* Cannot even emulate hcksum for unrecognized protocols */
489 	VIONA_PROBE2(fail_hcksum_proto, viona_link_t *, link, mblk_t *, mp);
490 	VIONA_RING_STAT_INCR(ring, fail_hcksum_proto);
491 	return (B_FALSE);
492 }
493 
494 static void
495 viona_tx(viona_link_t *link, viona_vring_t *ring)
496 {
497 	struct iovec		*iov = ring->vr_txiov;
498 	const uint_t		max_segs = ring->vr_size;
499 	uint16_t		cookie;
500 	int			i, n;
501 	uint32_t		len, base_off = 0;
502 	uint32_t		min_copy = VIONA_MAX_HDRS_LEN;
503 	mblk_t			*mp_head, *mp_tail, *mp;
504 	viona_desb_t		*dp = NULL;
505 	mac_client_handle_t	link_mch = link->l_mch;
506 	const struct virtio_net_hdr *hdr;
507 	vmm_page_t *pages = NULL;
508 
509 	mp_head = mp_tail = NULL;
510 
511 	ASSERT(iov != NULL);
512 
513 	n = vq_popchain(ring, iov, max_segs, &cookie, &pages);
514 	if (n == 0) {
515 		VIONA_PROBE1(tx_absent, viona_vring_t *, ring);
516 		VIONA_RING_STAT_INCR(ring, tx_absent);
517 		return;
518 	} else if (n < 0) {
519 		/*
520 		 * Any error encountered in vq_popchain has already resulted in
521 		 * specific probe and statistic handling.  Further action here
522 		 * is unnecessary.
523 		 */
524 		return;
525 	}
526 
527 	/* Grab the header and ensure it is of adequate length */
528 	hdr = (const struct virtio_net_hdr *)iov[0].iov_base;
529 	len = iov[0].iov_len;
530 	if (len < sizeof (struct virtio_net_hdr)) {
531 		goto drop_fail;
532 	}
533 
534 	/* Make sure the packet headers are always in the first mblk. */
535 	if (ring->vr_txdesb != NULL) {
536 		dp = &ring->vr_txdesb[cookie];
537 
538 		/*
539 		 * If the guest driver is operating properly, each desb slot
540 		 * should be available for use when processing a TX descriptor
541 		 * from the 'avail' ring.  In the case of drivers that reuse a
542 		 * descriptor before it has been posted to the 'used' ring, the
543 		 * data is simply dropped.
544 		 */
545 		if (atomic_cas_uint(&dp->d_ref, 0, 1) != 0) {
546 			dp = NULL;
547 			goto drop_fail;
548 		}
549 
550 		dp->d_cookie = cookie;
551 		mp_head = desballoc(dp->d_headers, VIONA_MAX_HDRS_LEN, 0,
552 		    &dp->d_frtn);
553 
554 		/* Account for the successful desballoc. */
555 		if (mp_head != NULL)
556 			dp->d_ref++;
557 	} else {
558 		mp_head = allocb(VIONA_MAX_HDRS_LEN, 0);
559 	}
560 
561 	if (mp_head == NULL)
562 		goto drop_fail;
563 
564 	mp_tail = mp_head;
565 
566 	/*
567 	 * We always copy enough of the guest data to cover the
568 	 * headers. This protects us from TOCTOU attacks and allows
569 	 * message block length assumptions to be made in subsequent
570 	 * code. In many cases, this means copying more data than
571 	 * strictly necessary. That's okay, as it is the larger packets
572 	 * (such as LSO) that really benefit from desballoc().
573 	 */
574 	for (i = 1; i < n; i++) {
575 		const uint32_t to_copy = MIN(min_copy, iov[i].iov_len);
576 
577 		bcopy(iov[i].iov_base, mp_head->b_wptr, to_copy);
578 		mp_head->b_wptr += to_copy;
579 		len += to_copy;
580 		min_copy -= to_copy;
581 
582 		/*
583 		 * We've met the minimum copy requirement. The rest of
584 		 * the guest data can be referenced.
585 		 */
586 		if (min_copy == 0) {
587 			/*
588 			 * If we copied all contents of this
589 			 * descriptor then move onto the next one.
590 			 * Otherwise, record how far we are into the
591 			 * current descriptor.
592 			 */
593 			if (iov[i].iov_len == to_copy)
594 				i++;
595 			else
596 				base_off = to_copy;
597 
598 			break;
599 		}
600 	}
601 
602 	ASSERT3P(mp_head, !=, NULL);
603 	ASSERT3P(mp_tail, !=, NULL);
604 
605 	for (; i < n; i++) {
606 		uintptr_t base = (uintptr_t)iov[i].iov_base + base_off;
607 		uint32_t chunk = iov[i].iov_len - base_off;
608 
609 		ASSERT3U(base_off, <, iov[i].iov_len);
610 		ASSERT3U(chunk, >, 0);
611 
612 		if (dp != NULL) {
613 			mp = desballoc((uchar_t *)base, chunk, 0, &dp->d_frtn);
614 			if (mp == NULL) {
615 				goto drop_fail;
616 			}
617 			dp->d_ref++;
618 		} else {
619 			mp = allocb(chunk, BPRI_MED);
620 			if (mp == NULL) {
621 				goto drop_fail;
622 			}
623 			bcopy((uchar_t *)base, mp->b_wptr, chunk);
624 		}
625 
626 		base_off = 0;
627 		len += chunk;
628 		mp->b_wptr += chunk;
629 		mp_tail->b_cont = mp;
630 		mp_tail = mp;
631 	}
632 
633 	if (VNETHOOK_INTERESTED_OUT(link->l_neti)) {
634 		/*
635 		 * The hook consumer may elect to free the mblk_t and set
636 		 * our mblk_t ** to NULL.  When using a viona_desb_t
637 		 * (dp != NULL), we do not want the corresponding cleanup to
638 		 * occur during the viona_hook() call. We instead want to
639 		 * reset and recycle dp for future use.  To prevent cleanup
640 		 * during the viona_hook() call, we take a ref on dp (if being
641 		 * used), and release it on success.  On failure, the
642 		 * freemsgchain() call will release all the refs taken earlier
643 		 * in viona_tx() (aside from the initial ref and the one we
644 		 * take), and drop_hook will reset dp for reuse.
645 		 */
646 		if (dp != NULL)
647 			dp->d_ref++;
648 
649 		/*
650 		 * Pass &mp instead of &mp_head so we don't lose track of
651 		 * mp_head if the hook consumer (i.e. ipf) elects to free mp
652 		 * and set mp to NULL.
653 		 */
654 		mp = mp_head;
655 		if (viona_hook(link, ring, &mp, B_TRUE) != 0) {
656 			if (mp != NULL)
657 				freemsgchain(mp);
658 			goto drop_hook;
659 		}
660 
661 		if (dp != NULL) {
662 			dp->d_ref--;
663 
664 			/*
665 			 * It is possible that the hook(s) accepted the packet,
666 			 * but as part of its processing, it issued a pull-up
667 			 * which released all references to the desb.  In that
668 			 * case, go back to acting like the packet is entirely
669 			 * copied (which it is).
670 			 */
671 			if (dp->d_ref == 1) {
672 				dp->d_cookie = 0;
673 				dp->d_ref = 0;
674 				dp = NULL;
675 			}
676 		}
677 	}
678 
679 	/*
680 	 * Request hardware checksumming, if necessary. If the guest
681 	 * sent an LSO packet then it must have also negotiated and
682 	 * requested partial checksum; therefore the LSO logic is
683 	 * contained within viona_tx_csum().
684 	 */
685 	if ((link->l_features & VIRTIO_NET_F_CSUM) != 0 &&
686 	    (hdr->vrh_flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) != 0) {
687 		if (!viona_tx_csum(ring, hdr, mp_head, len - iov[0].iov_len)) {
688 			goto drop_fail;
689 		}
690 	}
691 
692 	if (dp != NULL) {
693 		dp->d_len = len;
694 		dp->d_pages = pages;
695 		mutex_enter(&ring->vr_lock);
696 		ring->vr_xfer_outstanding++;
697 		mutex_exit(&ring->vr_lock);
698 	} else {
699 		/*
700 		 * If the data was cloned out of the ring, the descriptors can
701 		 * be marked as 'used' now, rather than deferring that action
702 		 * until after successful packet transmission.
703 		 */
704 		vmm_drv_page_release_chain(pages);
705 		viona_tx_done(ring, len, cookie);
706 	}
707 
708 	/*
709 	 * We're potentially going deep into the networking layer; make sure the
710 	 * guest can't run concurrently.
711 	 */
712 	smt_begin_unsafe();
713 	/*
714 	 * Ignore, for now, any signal from MAC about whether the outgoing
715 	 * packet was dropped or not.
716 	 */
717 	(void) mac_tx(link_mch, mp_head, 0, MAC_DROP_ON_NO_DESC, NULL);
718 	smt_end_unsafe();
719 	return;
720 
721 drop_fail:
722 	/*
723 	 * On the off chance that memory is not available via the desballoc or
724 	 * allocb calls, there are few options left besides to fail and drop
725 	 * the frame on the floor.
726 	 */
727 
728 	if (dp != NULL) {
729 		/*
730 		 * Take an additional reference on the desb handle (if present)
731 		 * so any desballoc-sourced mblks can release their hold on it
732 		 * without the handle reaching its final state and executing
733 		 * its clean-up logic.
734 		 */
735 		dp->d_ref++;
736 	}
737 
738 	/*
739 	 * Free any already-allocated blocks and sum up the total length of the
740 	 * dropped data to be released to the used ring.
741 	 */
742 	freemsgchain(mp_head);
743 
744 drop_hook:
745 	len = 0;
746 	for (uint_t i = 0; i < n; i++) {
747 		len += iov[i].iov_len;
748 	}
749 
750 	if (dp != NULL) {
751 		VERIFY(dp->d_ref == 2);
752 
753 		/* Clean up the desb handle, releasing the extra hold. */
754 		dp->d_len = 0;
755 		dp->d_cookie = 0;
756 		dp->d_ref = 0;
757 	}
758 
759 	VIONA_PROBE3(tx_drop, viona_vring_t *, ring, uint32_t, len,
760 	    uint16_t, cookie);
761 	vmm_drv_page_release_chain(pages);
762 	viona_tx_done(ring, len, cookie);
763 }
764