xref: /illumos-gate/usr/src/uts/intel/io/viona/viona_tx.c (revision 08855964b9970604433f7b19dcd71cf5af5e5f14)
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  * Tunable controls tx copy by default on or off
57  */
58 boolean_t viona_default_tx_copy = B_TRUE;
59 
60 /*
61  * copy tx mbufs from virtio ring to avoid necessitating a wait for packet
62  * transmission to free resources.
63  */
64 kmutex_t viona_force_copy_lock;
65 static enum viona_force_copy {
66 	VFC_UNINITALIZED	= 0,
67 	VFC_COPY_UNEEDED	= 1,
68 	VFC_COPY_REQUIRED	= 2,
69 } viona_force_copy_state = VFC_UNINITALIZED;
70 
71 struct viona_desb {
72 	frtn_t			d_frtn;
73 	viona_vring_t		*d_ring;
74 	uint_t			d_ref;
75 	uint32_t		d_len;
76 	uint16_t		d_cookie;
77 	uchar_t			*d_headers;
78 	vmm_page_t		*d_pages;
79 };
80 
81 static void viona_tx(viona_link_t *, viona_vring_t *);
82 static void viona_desb_release(viona_desb_t *);
83 
84 
85 static void
86 viona_tx_wait_outstanding(viona_vring_t *ring)
87 {
88 	ASSERT(MUTEX_HELD(&ring->vr_lock));
89 
90 	while (ring->vr_xfer_outstanding != 0) {
91 		/*
92 		 * Paying heed to signals is counterproductive here.  This is a
93 		 * very tight loop if pending transfers take an extended amount
94 		 * of time to be reclaimed while the host process is exiting.
95 		 */
96 		cv_wait(&ring->vr_cv, &ring->vr_lock);
97 	}
98 }
99 
100 /*
101  * Check if full TX packet copying is needed.  This should not be called from
102  * viona attach()/detach() context.
103  */
104 static boolean_t
105 viona_tx_copy_needed(void)
106 {
107 	boolean_t result;
108 
109 	if (viona_default_tx_copy) {
110 		return (B_TRUE);
111 	}
112 
113 	mutex_enter(&viona_force_copy_lock);
114 	if (viona_force_copy_state == VFC_UNINITALIZED) {
115 		major_t bnxe_major;
116 
117 		/*
118 		 * The original code for viona featured an explicit check for
119 		 * the bnxe driver which, when found present, necessitated that
120 		 * all transmissions be copied into their own mblks instead of
121 		 * passing guest memory to the underlying device.
122 		 *
123 		 * The motivations for this are unclear, but until it can be
124 		 * proven unnecessary, the check lives on.
125 		 */
126 		viona_force_copy_state = VFC_COPY_UNEEDED;
127 		if ((bnxe_major = ddi_name_to_major(BNXE_NIC_DRIVER))
128 		    != DDI_MAJOR_T_NONE) {
129 			if (ddi_hold_installed_driver(bnxe_major) != NULL) {
130 				viona_force_copy_state = VFC_COPY_REQUIRED;
131 				ddi_rele_driver(bnxe_major);
132 			}
133 		}
134 	}
135 	result = (viona_force_copy_state == VFC_COPY_REQUIRED);
136 	mutex_exit(&viona_force_copy_lock);
137 
138 	return (result);
139 }
140 
141 void
142 viona_tx_ring_alloc(viona_vring_t *ring, const uint16_t qsz)
143 {
144 	/* Allocate desb handles for TX ring if packet copying is disabled */
145 	if (!viona_tx_copy_needed()) {
146 		viona_desb_t *dp;
147 
148 		dp = kmem_zalloc(sizeof (viona_desb_t) * qsz, KM_SLEEP);
149 		ring->vr_txdesb = dp;
150 		for (uint_t i = 0; i < qsz; i++, dp++) {
151 			dp->d_frtn.free_func = viona_desb_release;
152 			dp->d_frtn.free_arg = (void *)dp;
153 			dp->d_ring = ring;
154 			dp->d_headers = kmem_zalloc(VIONA_MAX_HDRS_LEN,
155 			    KM_SLEEP);
156 		}
157 	}
158 
159 	/* Allocate ring-sized iovec buffers for TX */
160 	ring->vr_txiov = kmem_alloc(sizeof (struct iovec) * qsz, KM_SLEEP);
161 }
162 
163 void
164 viona_tx_ring_free(viona_vring_t *ring, const uint16_t qsz)
165 {
166 	if (ring->vr_txdesb != NULL) {
167 		viona_desb_t *dp = ring->vr_txdesb;
168 
169 		for (uint_t i = 0; i < qsz; i++, dp++) {
170 			kmem_free(dp->d_headers, VIONA_MAX_HDRS_LEN);
171 		}
172 		kmem_free(ring->vr_txdesb, sizeof (viona_desb_t) * qsz);
173 		ring->vr_txdesb = NULL;
174 	}
175 
176 	if (ring->vr_txiov != NULL) {
177 		kmem_free(ring->vr_txiov, sizeof (struct iovec) * qsz);
178 		ring->vr_txiov = NULL;
179 	}
180 }
181 
182 static void
183 viona_tx_done(viona_vring_t *ring, uint32_t len, uint16_t cookie)
184 {
185 	vq_pushchain(ring, len, cookie);
186 
187 	membar_enter();
188 	viona_intr_ring(ring, B_FALSE);
189 }
190 
191 #define	TX_BURST_THRESH	32
192 
193 void
194 viona_worker_tx(viona_vring_t *ring, viona_link_t *link)
195 {
196 	(void) thread_vsetname(curthread, "viona_tx_%p", ring);
197 
198 	ASSERT(MUTEX_HELD(&ring->vr_lock));
199 	ASSERT3U(ring->vr_state, ==, VRS_RUN);
200 
201 	mutex_exit(&ring->vr_lock);
202 
203 	for (;;) {
204 		uint_t ntx = 0, burst = 0;
205 
206 		viona_ring_disable_notify(ring);
207 		while (viona_ring_num_avail(ring) != 0) {
208 			viona_tx(link, ring);
209 			ntx++;
210 			burst++;
211 
212 			/*
213 			 * It is advantageous for throughput to keep this
214 			 * transmission loop tight, but periodic breaks to
215 			 * check for other events are of value too.
216 			 */
217 			if (burst >= TX_BURST_THRESH) {
218 				mutex_enter(&ring->vr_lock);
219 				const bool need_bail = vring_need_bail(ring);
220 				mutex_exit(&ring->vr_lock);
221 
222 				if (need_bail) {
223 					break;
224 				}
225 				burst = 0;
226 			}
227 		}
228 
229 		VIONA_PROBE2(tx, viona_link_t *, link, uint_t, ntx);
230 
231 		/*
232 		 * Check for available descriptors on the ring once more in
233 		 * case a late addition raced with the NO_NOTIFY flag toggle.
234 		 *
235 		 * The barrier ensures that visibility of the no-notify
236 		 * store does not cross the viona_ring_num_avail() check below.
237 		 */
238 		viona_ring_enable_notify(ring);
239 		membar_enter();
240 
241 		if (viona_ring_num_avail(ring) == 0 &&
242 		    (link->l_features & VIRTIO_F_RING_NOTIFY_ON_EMPTY) != 0) {
243 			/*
244 			 * The NOTIFY_ON_EMPTY interrupt should not pay heed to
245 			 * the presence of AVAIL_NO_INTERRUPT.
246 			 */
247 			viona_intr_ring(ring, B_TRUE);
248 		}
249 
250 		mutex_enter(&ring->vr_lock);
251 		for (;;) {
252 			if (vring_need_bail(ring)) {
253 				ring->vr_state = VRS_STOP;
254 				viona_tx_wait_outstanding(ring);
255 				return;
256 			}
257 
258 			if (vmm_drv_lease_expired(ring->vr_lease)) {
259 				ring->vr_state_flags |= VRSF_RENEW;
260 				/*
261 				 * When renewing the lease for the ring, no TX
262 				 * frames may be outstanding, as they contain
263 				 * references to guest memory.
264 				 */
265 				viona_tx_wait_outstanding(ring);
266 
267 				const boolean_t renewed =
268 				    viona_ring_lease_renew(ring);
269 				ring->vr_state_flags &= ~VRSF_RENEW;
270 
271 				if (!renewed) {
272 					/* stop ring on failed renewal */
273 					ring->vr_state = VRS_STOP;
274 					return;
275 				}
276 			}
277 
278 			if (viona_ring_num_avail(ring) != 0) {
279 				break;
280 			}
281 
282 			/* Wait for further activity on the ring */
283 			(void) cv_wait_sig(&ring->vr_cv, &ring->vr_lock);
284 		}
285 		mutex_exit(&ring->vr_lock);
286 	}
287 	/* UNREACHABLE */
288 }
289 
290 static void
291 viona_desb_release(viona_desb_t *dp)
292 {
293 	viona_vring_t *ring = dp->d_ring;
294 	uint_t ref;
295 	uint32_t len;
296 	uint16_t cookie;
297 
298 	ref = atomic_dec_uint_nv(&dp->d_ref);
299 	if (ref > 1) {
300 		return;
301 	}
302 
303 	/*
304 	 * The desb corresponding to this index must be ready for reuse before
305 	 * the descriptor is returned to the guest via the 'used' ring.
306 	 */
307 	len = dp->d_len;
308 	cookie = dp->d_cookie;
309 	dp->d_len = 0;
310 	dp->d_cookie = 0;
311 	vmm_drv_page_release_chain(dp->d_pages);
312 	dp->d_pages = NULL;
313 
314 	/*
315 	 * Ensure all other changes to the desb are visible prior to zeroing its
316 	 * refcount, signifying its readiness for reuse.
317 	 */
318 	membar_exit();
319 	dp->d_ref = 0;
320 
321 	viona_tx_done(ring, len, cookie);
322 
323 	mutex_enter(&ring->vr_lock);
324 	if ((--ring->vr_xfer_outstanding) == 0) {
325 		cv_broadcast(&ring->vr_cv);
326 	}
327 	mutex_exit(&ring->vr_lock);
328 }
329 
330 static boolean_t
331 viona_tx_csum(viona_vring_t *ring, const struct virtio_net_hdr *hdr,
332     mblk_t *mp, uint32_t len)
333 {
334 	viona_link_t *link = ring->vr_link;
335 	const struct ether_header *eth;
336 	uint_t eth_len = sizeof (struct ether_header);
337 	ushort_t ftype;
338 	ipha_t *ipha = NULL;
339 	uint8_t ipproto = IPPROTO_NONE; /* NONE is not exactly right, but ok */
340 	uint16_t flags = 0;
341 	const uint_t csum_start = hdr->vrh_csum_start;
342 	const uint_t csum_stuff = hdr->vrh_csum_offset + csum_start;
343 
344 	/*
345 	 * Validate that the checksum offsets provided by the guest are within
346 	 * the bounds of the packet.  Additionally, ensure that the checksum
347 	 * contents field is within the headers mblk copied by viona_tx().
348 	 */
349 	if (csum_start >= len || csum_start < eth_len || csum_stuff >= len ||
350 	    (csum_stuff + sizeof (uint16_t)) > MBLKL(mp)) {
351 		VIONA_PROBE2(fail_hcksum, viona_link_t *, link, mblk_t *, mp);
352 		VIONA_RING_STAT_INCR(ring, fail_hcksum);
353 		return (B_FALSE);
354 	}
355 
356 	/*
357 	 * This is guaranteed to be safe thanks to the header copying
358 	 * done in viona_tx().
359 	 */
360 	eth = (const struct ether_header *)mp->b_rptr;
361 	ftype = ntohs(eth->ether_type);
362 
363 	if (ftype == ETHERTYPE_VLAN) {
364 		const struct ether_vlan_header *veth;
365 
366 		/* punt on QinQ for now */
367 		eth_len = sizeof (struct ether_vlan_header);
368 		veth = (const struct ether_vlan_header *)eth;
369 		ftype = ntohs(veth->ether_type);
370 	}
371 
372 	if (ftype == ETHERTYPE_IP) {
373 		ipha = (ipha_t *)(mp->b_rptr + eth_len);
374 
375 		ipproto = ipha->ipha_protocol;
376 	} else if (ftype == ETHERTYPE_IPV6) {
377 		ip6_t *ip6h = (ip6_t *)(mp->b_rptr + eth_len);
378 
379 		ipproto = ip6h->ip6_nxt;
380 	}
381 
382 	/*
383 	 * We ignore hdr_len because the spec says it can't be
384 	 * trusted. Besides, our own stack will determine the header
385 	 * boundary.
386 	 */
387 	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0 &&
388 	    (hdr->vrh_gso_type & VIRTIO_NET_HDR_GSO_TCPV4) != 0 &&
389 	    ftype == ETHERTYPE_IP) {
390 		uint16_t	*cksump;
391 		uint32_t	cksum;
392 		ipaddr_t	src = ipha->ipha_src;
393 		ipaddr_t	dst = ipha->ipha_dst;
394 
395 		/*
396 		 * Our native IP stack doesn't set the L4 length field
397 		 * of the pseudo header when LSO is in play. Other IP
398 		 * stacks, e.g. Linux, do include the length field.
399 		 * This is a problem because the hardware expects that
400 		 * the length field is not set. When it is set it will
401 		 * cause an incorrect TCP checksum to be generated.
402 		 * The reason this works in Linux is because Linux
403 		 * corrects the pseudo-header checksum in the driver
404 		 * code. In order to get the correct HW checksum we
405 		 * need to assume the guest's IP stack gave us a bogus
406 		 * TCP partial checksum and calculate it ourselves.
407 		 */
408 		cksump = IPH_TCPH_CHECKSUMP(ipha, IPH_HDR_LENGTH(ipha));
409 		cksum = IP_TCP_CSUM_COMP;
410 		cksum += (dst >> 16) + (dst & 0xFFFF) +
411 		    (src >> 16) + (src & 0xFFFF);
412 		cksum = (cksum & 0xFFFF) + (cksum >> 16);
413 		*(cksump) = (cksum & 0xFFFF) + (cksum >> 16);
414 
415 		/*
416 		 * Since viona is a "legacy device", the data stored
417 		 * by the driver will be in the guest's native endian
418 		 * format (see sections 2.4.3 and 5.1.6.1 of the
419 		 * VIRTIO 1.0 spec for more info). At this time the
420 		 * only guests using viona are x86 and we can assume
421 		 * little-endian.
422 		 */
423 		lso_info_set(mp, LE_16(hdr->vrh_gso_size), HW_LSO);
424 
425 		/*
426 		 * Hardware, like ixgbe, expects the client to request
427 		 * IP header checksum offload if it's sending LSO (see
428 		 * ixgbe_get_context()). Unfortunately, virtio makes
429 		 * no allowances for negotiating IP header checksum
430 		 * and HW offload, only TCP checksum. We add the flag
431 		 * and zero-out the checksum field. This mirrors the
432 		 * behavior of our native IP stack (which does this in
433 		 * the interest of HW that expects the field to be
434 		 * zero).
435 		 */
436 		flags |= HCK_IPV4_HDRCKSUM;
437 		ipha->ipha_hdr_checksum = 0;
438 	}
439 
440 	/*
441 	 * Use DB_CKSUMFLAGS instead of mac_hcksum_get() to make sure
442 	 * HW_LSO, if present, is not lost.
443 	 */
444 	flags |= DB_CKSUMFLAGS(mp);
445 
446 	/*
447 	 * Partial checksum support from the NIC is ideal, since it most
448 	 * closely maps to the interface defined by virtio.
449 	 */
450 	if ((link->l_cap_csum & HCKSUM_INET_PARTIAL) != 0 &&
451 	    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
452 		/*
453 		 * MAC expects these offsets to be relative to the
454 		 * start of the L3 header rather than the L2 frame.
455 		 */
456 		flags |= HCK_PARTIALCKSUM;
457 		mac_hcksum_set(mp, csum_start - eth_len, csum_stuff - eth_len,
458 		    len - eth_len, 0, flags);
459 		return (B_TRUE);
460 	}
461 
462 	/*
463 	 * Without partial checksum support, look to the L3/L4 protocol
464 	 * information to see if the NIC can handle it.  If not, the
465 	 * checksum will need to calculated inline.
466 	 */
467 	if (ftype == ETHERTYPE_IP) {
468 		if ((link->l_cap_csum & HCKSUM_INET_FULL_V4) != 0 &&
469 		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
470 			uint16_t *csump = (uint16_t *)(mp->b_rptr + csum_stuff);
471 			*csump = 0;
472 			flags |= HCK_FULLCKSUM;
473 			mac_hcksum_set(mp, 0, 0, 0, 0, flags);
474 			return (B_TRUE);
475 		}
476 
477 		/* XXX: Implement manual fallback checksumming? */
478 		VIONA_PROBE2(fail_hcksum, viona_link_t *, link, mblk_t *, mp);
479 		VIONA_RING_STAT_INCR(ring, fail_hcksum);
480 		return (B_FALSE);
481 	} else if (ftype == ETHERTYPE_IPV6) {
482 		if ((link->l_cap_csum & HCKSUM_INET_FULL_V6) != 0 &&
483 		    (ipproto == IPPROTO_TCP || ipproto == IPPROTO_UDP)) {
484 			uint16_t *csump = (uint16_t *)(mp->b_rptr + csum_stuff);
485 			*csump = 0;
486 			flags |= HCK_FULLCKSUM;
487 			mac_hcksum_set(mp, 0, 0, 0, 0, flags);
488 			return (B_TRUE);
489 		}
490 
491 		/* XXX: Implement manual fallback checksumming? */
492 		VIONA_PROBE2(fail_hcksum6, viona_link_t *, link, mblk_t *, mp);
493 		VIONA_RING_STAT_INCR(ring, fail_hcksum6);
494 		return (B_FALSE);
495 	}
496 
497 	/* Cannot even emulate hcksum for unrecognized protocols */
498 	VIONA_PROBE2(fail_hcksum_proto, viona_link_t *, link, mblk_t *, mp);
499 	VIONA_RING_STAT_INCR(ring, fail_hcksum_proto);
500 	return (B_FALSE);
501 }
502 
503 static void
504 viona_tx(viona_link_t *link, viona_vring_t *ring)
505 {
506 	struct iovec		*iov = ring->vr_txiov;
507 	const uint_t		max_segs = ring->vr_size;
508 	uint16_t		cookie;
509 	int			i, n;
510 	uint32_t		len, base_off = 0;
511 	uint32_t		min_copy = VIONA_MAX_HDRS_LEN;
512 	mblk_t			*mp_head, *mp_tail, *mp;
513 	viona_desb_t		*dp = NULL;
514 	mac_client_handle_t	link_mch = link->l_mch;
515 	const struct virtio_net_hdr *hdr;
516 	vmm_page_t *pages = NULL;
517 
518 	mp_head = mp_tail = NULL;
519 
520 	ASSERT(iov != NULL);
521 
522 	n = vq_popchain(ring, iov, max_segs, &cookie, &pages);
523 	if (n == 0) {
524 		VIONA_PROBE1(tx_absent, viona_vring_t *, ring);
525 		VIONA_RING_STAT_INCR(ring, tx_absent);
526 		return;
527 	} else if (n < 0) {
528 		/*
529 		 * Any error encountered in vq_popchain has already resulted in
530 		 * specific probe and statistic handling.  Further action here
531 		 * is unnecessary.
532 		 */
533 		return;
534 	}
535 
536 	/* Grab the header and ensure it is of adequate length */
537 	hdr = (const struct virtio_net_hdr *)iov[0].iov_base;
538 	len = iov[0].iov_len;
539 	if (len < sizeof (struct virtio_net_hdr)) {
540 		goto drop_fail;
541 	}
542 
543 	/* Make sure the packet headers are always in the first mblk. */
544 	if (ring->vr_txdesb != NULL) {
545 		dp = &ring->vr_txdesb[cookie];
546 
547 		/*
548 		 * If the guest driver is operating properly, each desb slot
549 		 * should be available for use when processing a TX descriptor
550 		 * from the 'avail' ring.  In the case of drivers that reuse a
551 		 * descriptor before it has been posted to the 'used' ring, the
552 		 * data is simply dropped.
553 		 */
554 		if (atomic_cas_uint(&dp->d_ref, 0, 1) != 0) {
555 			dp = NULL;
556 			goto drop_fail;
557 		}
558 
559 		dp->d_cookie = cookie;
560 		mp_head = desballoc(dp->d_headers, VIONA_MAX_HDRS_LEN, 0,
561 		    &dp->d_frtn);
562 
563 		/* Account for the successful desballoc. */
564 		if (mp_head != NULL)
565 			dp->d_ref++;
566 	} else {
567 		mp_head = allocb(VIONA_MAX_HDRS_LEN, 0);
568 	}
569 
570 	if (mp_head == NULL)
571 		goto drop_fail;
572 
573 	mp_tail = mp_head;
574 
575 	/*
576 	 * We always copy enough of the guest data to cover the
577 	 * headers. This protects us from TOCTOU attacks and allows
578 	 * message block length assumptions to be made in subsequent
579 	 * code. In many cases, this means copying more data than
580 	 * strictly necessary. That's okay, as it is the larger packets
581 	 * (such as LSO) that really benefit from desballoc().
582 	 */
583 	for (i = 1; i < n; i++) {
584 		const uint32_t to_copy = MIN(min_copy, iov[i].iov_len);
585 
586 		bcopy(iov[i].iov_base, mp_head->b_wptr, to_copy);
587 		mp_head->b_wptr += to_copy;
588 		len += to_copy;
589 		min_copy -= to_copy;
590 
591 		/*
592 		 * We've met the minimum copy requirement. The rest of
593 		 * the guest data can be referenced.
594 		 */
595 		if (min_copy == 0) {
596 			/*
597 			 * If we copied all contents of this
598 			 * descriptor then move onto the next one.
599 			 * Otherwise, record how far we are into the
600 			 * current descriptor.
601 			 */
602 			if (iov[i].iov_len == to_copy)
603 				i++;
604 			else
605 				base_off = to_copy;
606 
607 			break;
608 		}
609 	}
610 
611 	ASSERT3P(mp_head, !=, NULL);
612 	ASSERT3P(mp_tail, !=, NULL);
613 
614 	for (; i < n; i++) {
615 		uintptr_t base = (uintptr_t)iov[i].iov_base + base_off;
616 		uint32_t chunk = iov[i].iov_len - base_off;
617 
618 		ASSERT3U(base_off, <, iov[i].iov_len);
619 		ASSERT3U(chunk, >, 0);
620 
621 		if (dp != NULL) {
622 			mp = desballoc((uchar_t *)base, chunk, 0, &dp->d_frtn);
623 			if (mp == NULL) {
624 				goto drop_fail;
625 			}
626 			dp->d_ref++;
627 		} else {
628 			mp = allocb(chunk, BPRI_MED);
629 			if (mp == NULL) {
630 				goto drop_fail;
631 			}
632 			bcopy((uchar_t *)base, mp->b_wptr, chunk);
633 		}
634 
635 		base_off = 0;
636 		len += chunk;
637 		mp->b_wptr += chunk;
638 		mp_tail->b_cont = mp;
639 		mp_tail = mp;
640 	}
641 
642 	if (VNETHOOK_INTERESTED_OUT(link->l_neti)) {
643 		/*
644 		 * The hook consumer may elect to free the mblk_t and set
645 		 * our mblk_t ** to NULL.  When using a viona_desb_t
646 		 * (dp != NULL), we do not want the corresponding cleanup to
647 		 * occur during the viona_hook() call. We instead want to
648 		 * reset and recycle dp for future use.  To prevent cleanup
649 		 * during the viona_hook() call, we take a ref on dp (if being
650 		 * used), and release it on success.  On failure, the
651 		 * freemsgchain() call will release all the refs taken earlier
652 		 * in viona_tx() (aside from the initial ref and the one we
653 		 * take), and drop_hook will reset dp for reuse.
654 		 */
655 		if (dp != NULL)
656 			dp->d_ref++;
657 
658 		/*
659 		 * Pass &mp instead of &mp_head so we don't lose track of
660 		 * mp_head if the hook consumer (i.e. ipf) elects to free mp
661 		 * and set mp to NULL.
662 		 */
663 		mp = mp_head;
664 		if (viona_hook(link, ring, &mp, B_TRUE) != 0) {
665 			if (mp != NULL)
666 				freemsgchain(mp);
667 			goto drop_hook;
668 		}
669 
670 		if (dp != NULL) {
671 			dp->d_ref--;
672 
673 			/*
674 			 * It is possible that the hook(s) accepted the packet,
675 			 * but as part of its processing, it issued a pull-up
676 			 * which released all references to the desb.  In that
677 			 * case, go back to acting like the packet is entirely
678 			 * copied (which it is).
679 			 */
680 			if (dp->d_ref == 1) {
681 				dp->d_cookie = 0;
682 				dp->d_ref = 0;
683 				dp = NULL;
684 			}
685 		}
686 	}
687 
688 	/*
689 	 * Request hardware checksumming, if necessary. If the guest
690 	 * sent an LSO packet then it must have also negotiated and
691 	 * requested partial checksum; therefore the LSO logic is
692 	 * contained within viona_tx_csum().
693 	 */
694 	if ((link->l_features & VIRTIO_NET_F_CSUM) != 0 &&
695 	    (hdr->vrh_flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) != 0) {
696 		if (!viona_tx_csum(ring, hdr, mp_head, len - iov[0].iov_len)) {
697 			goto drop_fail;
698 		}
699 	}
700 
701 	if (dp != NULL) {
702 		dp->d_len = len;
703 		dp->d_pages = pages;
704 		mutex_enter(&ring->vr_lock);
705 		ring->vr_xfer_outstanding++;
706 		mutex_exit(&ring->vr_lock);
707 	} else {
708 		/*
709 		 * If the data was cloned out of the ring, the descriptors can
710 		 * be marked as 'used' now, rather than deferring that action
711 		 * until after successful packet transmission.
712 		 */
713 		vmm_drv_page_release_chain(pages);
714 		viona_tx_done(ring, len, cookie);
715 	}
716 
717 	/*
718 	 * We're potentially going deep into the networking layer; make sure the
719 	 * guest can't run concurrently.
720 	 */
721 	smt_begin_unsafe();
722 	/*
723 	 * Ignore, for now, any signal from MAC about whether the outgoing
724 	 * packet was dropped or not.
725 	 */
726 	(void) mac_tx(link_mch, mp_head, 0, MAC_DROP_ON_NO_DESC, NULL);
727 	smt_end_unsafe();
728 	return;
729 
730 drop_fail:
731 	/*
732 	 * On the off chance that memory is not available via the desballoc or
733 	 * allocb calls, there are few options left besides to fail and drop
734 	 * the frame on the floor.
735 	 */
736 
737 	if (dp != NULL) {
738 		/*
739 		 * Take an additional reference on the desb handle (if present)
740 		 * so any desballoc-sourced mblks can release their hold on it
741 		 * without the handle reaching its final state and executing
742 		 * its clean-up logic.
743 		 */
744 		dp->d_ref++;
745 	}
746 
747 	/*
748 	 * Free any already-allocated blocks and sum up the total length of the
749 	 * dropped data to be released to the used ring.
750 	 */
751 	freemsgchain(mp_head);
752 
753 drop_hook:
754 	len = 0;
755 	for (uint_t i = 0; i < n; i++) {
756 		len += iov[i].iov_len;
757 	}
758 
759 	if (dp != NULL) {
760 		VERIFY(dp->d_ref == 2);
761 
762 		/* Clean up the desb handle, releasing the extra hold. */
763 		dp->d_len = 0;
764 		dp->d_cookie = 0;
765 		dp->d_ref = 0;
766 	}
767 
768 	VIONA_PROBE3(tx_drop, viona_vring_t *, ring, uint32_t, len,
769 	    uint16_t, cookie);
770 	vmm_drv_page_release_chain(pages);
771 	viona_tx_done(ring, len, cookie);
772 }
773