xref: /illumos-gate/usr/src/uts/common/io/nxge/nxge_send.c (revision 2264ca7f5db194583c672cb5779a67f52bcd92a9)
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  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/mac_provider.h>
27 #include <sys/nxge/nxge_impl.h>
28 #include <sys/nxge/nxge_hio.h>
29 #include <npi_tx_wr64.h>
30 
31 /* Software LSO required header files */
32 #include <netinet/tcp.h>
33 #include <inet/ip_impl.h>
34 #include <inet/tcp.h>
35 
36 extern uint64_t mac_pkt_hash(uint_t, mblk_t *mp, uint8_t policy,
37     boolean_t is_outbound);
38 
39 static mblk_t *nxge_lso_eliminate(mblk_t *);
40 static mblk_t *nxge_do_softlso(mblk_t *mp, uint32_t mss);
41 static void nxge_lso_info_get(mblk_t *, uint32_t *, uint32_t *);
42 static void nxge_hcksum_retrieve(mblk_t *,
43     uint32_t *, uint32_t *, uint32_t *,
44     uint32_t *, uint32_t *);
45 static uint32_t nxge_csgen(uint16_t *, int);
46 
47 extern uint32_t		nxge_reclaim_pending;
48 extern uint32_t 	nxge_bcopy_thresh;
49 extern uint32_t 	nxge_dvma_thresh;
50 extern uint32_t 	nxge_dma_stream_thresh;
51 extern uint32_t		nxge_tx_minfree;
52 extern uint32_t		nxge_tx_intr_thres;
53 extern uint32_t		nxge_tx_max_gathers;
54 extern uint32_t		nxge_tx_tiny_pack;
55 extern uint32_t		nxge_tx_use_bcopy;
56 extern nxge_tx_mode_t	nxge_tx_scheme;
57 uint32_t		nxge_lso_kick_cnt = 2;
58 
59 
60 void
61 nxge_tx_ring_task(void *arg)
62 {
63 	p_tx_ring_t	ring = (p_tx_ring_t)arg;
64 
65 	MUTEX_ENTER(&ring->lock);
66 	(void) nxge_txdma_reclaim(ring->nxgep, ring, 0);
67 	MUTEX_EXIT(&ring->lock);
68 
69 	if (!isLDOMguest(ring->nxgep) && !ring->tx_ring_offline)
70 		mac_tx_ring_update(ring->nxgep->mach, ring->tx_ring_handle);
71 #if defined(sun4v)
72 	else {
73 		nxge_hio_data_t *nhd =
74 		    (nxge_hio_data_t *)ring->nxgep->nxge_hw_p->hio;
75 		nx_vio_fp_t *vio = &nhd->hio.vio;
76 
77 		/* Call back vnet. */
78 		if (vio->cb.vio_net_tx_update) {
79 			(*vio->cb.vio_net_tx_update)(ring->nxgep->hio_vr->vhp);
80 		}
81 	}
82 #endif
83 }
84 
85 static void
86 nxge_tx_ring_dispatch(p_tx_ring_t ring)
87 {
88 	/*
89 	 * Kick the ring task to reclaim some buffers.
90 	 */
91 	(void) ddi_taskq_dispatch(ring->taskq,
92 	    nxge_tx_ring_task, (void *)ring, DDI_SLEEP);
93 }
94 
95 mblk_t *
96 nxge_tx_ring_send(void *arg, mblk_t *mp)
97 {
98 	p_nxge_ring_handle_t	nrhp = (p_nxge_ring_handle_t)arg;
99 	p_nxge_t		nxgep;
100 	p_tx_ring_t		tx_ring_p;
101 	int			status, channel;
102 
103 	ASSERT(nrhp != NULL);
104 	nxgep = nrhp->nxgep;
105 	channel = nxgep->pt_config.hw_config.tdc.start + nrhp->index;
106 	tx_ring_p = nxgep->tx_rings->rings[channel];
107 
108 	ASSERT(nxgep == tx_ring_p->nxgep);
109 
110 #ifdef DEBUG
111 	if (isLDOMservice(nxgep)) {
112 		ASSERT(!tx_ring_p->tx_ring_offline);
113 	}
114 #endif
115 
116 	status = nxge_start(nxgep, tx_ring_p, mp);
117 	if (status) {
118 		nxge_tx_ring_dispatch(tx_ring_p);
119 		return (mp);
120 	}
121 
122 	return ((mblk_t *)NULL);
123 }
124 
125 #if defined(sun4v)
126 
127 /*
128  * Hashing policy for load balancing over the set of TX rings
129  * available to the driver.
130  */
131 static uint8_t nxge_tx_hash_policy = MAC_PKT_HASH_L4;
132 
133 /*
134  * nxge_m_tx() is needed for Hybrid I/O operation of the vnet in
135  *	the guest domain.  See CR 6778758 for long term solution.
136  *
137  *	The guest domain driver will for now hash the packet
138  *	to pick a DMA channel from the only group it has group 0.
139  */
140 
141 mblk_t *
142 nxge_m_tx(void *arg, mblk_t *mp)
143 {
144 	p_nxge_t		nxgep = (p_nxge_t)arg;
145 	mblk_t			*next;
146 	uint64_t		rindex;
147 	p_tx_ring_t		tx_ring_p;
148 	int			status;
149 
150 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_m_tx"));
151 
152 	/*
153 	 * Hash to pick a ring from Group 0, the only TX group
154 	 * for a guest domain driver.
155 	 */
156 	rindex = mac_pkt_hash(DL_ETHER, mp, nxge_tx_hash_policy, B_TRUE);
157 	rindex = rindex % nxgep->pt_config.tdc_grps[0].max_tdcs;
158 
159 	/*
160 	 * Get the ring handle.
161 	 */
162 	tx_ring_p = nxgep->tx_rings->rings[rindex];
163 
164 	while (mp != NULL) {
165 		next = mp->b_next;
166 		mp->b_next = NULL;
167 
168 		status = nxge_start(nxgep, tx_ring_p, mp);
169 		if (status != 0) {
170 			mp->b_next = next;
171 			nxge_tx_ring_dispatch(tx_ring_p);
172 			return (mp);
173 		}
174 
175 		mp = next;
176 	}
177 
178 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_m_tx"));
179 	return ((mblk_t *)NULL);
180 }
181 
182 #endif
183 
184 int
185 nxge_start(p_nxge_t nxgep, p_tx_ring_t tx_ring_p, p_mblk_t mp)
186 {
187 	int 			status = 0;
188 	p_tx_desc_t 		tx_desc_ring_vp;
189 	npi_handle_t		npi_desc_handle;
190 	nxge_os_dma_handle_t 	tx_desc_dma_handle;
191 	p_tx_desc_t 		tx_desc_p;
192 	p_tx_msg_t 		tx_msg_ring;
193 	p_tx_msg_t 		tx_msg_p;
194 	tx_desc_t		tx_desc, *tmp_desc_p;
195 	tx_desc_t		sop_tx_desc, *sop_tx_desc_p;
196 	p_tx_pkt_header_t	hdrp;
197 	tx_pkt_hdr_all_t	tmp_hdrp;
198 	p_tx_pkt_hdr_all_t	pkthdrp;
199 	uint8_t			npads = 0;
200 	uint64_t 		dma_ioaddr;
201 	uint32_t		dma_flags;
202 	int			last_bidx;
203 	uint8_t 		*b_rptr;
204 	caddr_t 		kaddr;
205 	uint32_t		nmblks;
206 	uint32_t		ngathers;
207 	uint32_t		clen;
208 	int 			len;
209 	uint32_t		pkt_len, pack_len, min_len;
210 	uint32_t		bcopy_thresh;
211 	int 			i, cur_index, sop_index;
212 	uint16_t		tail_index;
213 	boolean_t		tail_wrap = B_FALSE;
214 	nxge_dma_common_t	desc_area;
215 	nxge_os_dma_handle_t 	dma_handle;
216 	ddi_dma_cookie_t 	dma_cookie;
217 	npi_handle_t		npi_handle;
218 	p_mblk_t 		nmp;
219 	p_mblk_t		t_mp;
220 	uint32_t 		ncookies;
221 	boolean_t 		good_packet;
222 	boolean_t 		mark_mode = B_FALSE;
223 	p_nxge_stats_t 		statsp;
224 	p_nxge_tx_ring_stats_t tdc_stats;
225 	t_uscalar_t 		start_offset = 0;
226 	t_uscalar_t 		stuff_offset = 0;
227 	t_uscalar_t 		end_offset = 0;
228 	t_uscalar_t 		value = 0;
229 	t_uscalar_t 		cksum_flags = 0;
230 	boolean_t		cksum_on = B_FALSE;
231 	uint32_t		boff = 0;
232 	uint64_t		tot_xfer_len = 0;
233 	boolean_t		header_set = B_FALSE;
234 #ifdef NXGE_DEBUG
235 	p_tx_desc_t 		tx_desc_ring_pp;
236 	p_tx_desc_t 		tx_desc_pp;
237 	tx_desc_t		*save_desc_p;
238 	int			dump_len;
239 	int			sad_len;
240 	uint64_t		sad;
241 	int			xfer_len;
242 	uint32_t		msgsize;
243 #endif
244 	p_mblk_t 		mp_chain = NULL;
245 	boolean_t		is_lso = B_FALSE;
246 	boolean_t		lso_again;
247 	int			cur_index_lso;
248 	p_mblk_t 		nmp_lso_save;
249 	uint32_t		lso_ngathers;
250 	boolean_t		lso_tail_wrap = B_FALSE;
251 
252 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
253 	    "==> nxge_start: tx dma channel %d", tx_ring_p->tdc));
254 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
255 	    "==> nxge_start: Starting tdc %d desc pending %d",
256 	    tx_ring_p->tdc, tx_ring_p->descs_pending));
257 
258 	statsp = nxgep->statsp;
259 
260 	if (!isLDOMguest(nxgep)) {
261 		switch (nxgep->mac.portmode) {
262 		default:
263 			if (nxgep->statsp->port_stats.lb_mode ==
264 			    nxge_lb_normal) {
265 				if (!statsp->mac_stats.link_up) {
266 					freemsg(mp);
267 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
268 					    "==> nxge_start: "
269 					    "link not up"));
270 					goto nxge_start_fail1;
271 				}
272 			}
273 			break;
274 		case PORT_10G_FIBER:
275 			/*
276 			 * For the following modes, check the link status
277 			 * before sending the packet out:
278 			 * nxge_lb_normal,
279 			 * nxge_lb_ext10g,
280 			 * nxge_lb_ext1000,
281 			 * nxge_lb_ext100,
282 			 * nxge_lb_ext10.
283 			 */
284 			if (nxgep->statsp->port_stats.lb_mode <
285 			    nxge_lb_phy10g) {
286 				if (!statsp->mac_stats.link_up) {
287 					freemsg(mp);
288 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
289 					    "==> nxge_start: "
290 					    "link not up"));
291 					goto nxge_start_fail1;
292 				}
293 			}
294 			break;
295 		}
296 	}
297 
298 	if ((!(nxgep->drv_state & STATE_HW_INITIALIZED)) ||
299 	    (nxgep->nxge_mac_state != NXGE_MAC_STARTED)) {
300 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
301 		    "==> nxge_start: hardware not initialized or stopped"));
302 		freemsg(mp);
303 		goto nxge_start_fail1;
304 	}
305 
306 	if (nxgep->soft_lso_enable) {
307 		mp_chain = nxge_lso_eliminate(mp);
308 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
309 		    "==> nxge_start(0): LSO mp $%p mp_chain $%p",
310 		    mp, mp_chain));
311 		if (mp_chain == NULL) {
312 			NXGE_ERROR_MSG((nxgep, TX_CTL,
313 			    "==> nxge_send(0): NULL mp_chain $%p != mp $%p",
314 			    mp_chain, mp));
315 			goto nxge_start_fail1;
316 		}
317 		if (mp_chain != mp) {
318 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
319 			    "==> nxge_send(1): IS LSO mp_chain $%p != mp $%p",
320 			    mp_chain, mp));
321 			is_lso = B_TRUE;
322 			mp = mp_chain;
323 			mp_chain = mp_chain->b_next;
324 			mp->b_next = NULL;
325 		}
326 	}
327 
328 	hcksum_retrieve(mp, NULL, NULL, &start_offset,
329 	    &stuff_offset, &end_offset, &value, &cksum_flags);
330 	if (!NXGE_IS_VLAN_PACKET(mp->b_rptr)) {
331 		start_offset += sizeof (ether_header_t);
332 		stuff_offset += sizeof (ether_header_t);
333 	} else {
334 		start_offset += sizeof (struct ether_vlan_header);
335 		stuff_offset += sizeof (struct ether_vlan_header);
336 	}
337 
338 	if (cksum_flags & HCK_PARTIALCKSUM) {
339 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
340 		    "==> nxge_start: mp $%p len %d "
341 		    "cksum_flags 0x%x (partial checksum) ",
342 		    mp, MBLKL(mp), cksum_flags));
343 		cksum_on = B_TRUE;
344 	}
345 
346 	pkthdrp = (p_tx_pkt_hdr_all_t)&tmp_hdrp;
347 	pkthdrp->reserved = 0;
348 	tmp_hdrp.pkthdr.value = 0;
349 	nxge_fill_tx_hdr(mp, B_FALSE, cksum_on,
350 	    0, 0, pkthdrp,
351 	    start_offset, stuff_offset);
352 
353 	lso_again = B_FALSE;
354 	lso_ngathers = 0;
355 
356 	MUTEX_ENTER(&tx_ring_p->lock);
357 
358 	if (isLDOMservice(nxgep)) {
359 		tx_ring_p->tx_ring_busy = B_TRUE;
360 		if (tx_ring_p->tx_ring_offline) {
361 			freemsg(mp);
362 			tx_ring_p->tx_ring_busy = B_FALSE;
363 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
364 			    NXGE_TX_RING_OFFLINED);
365 			MUTEX_EXIT(&tx_ring_p->lock);
366 			return (status);
367 		}
368 	}
369 
370 	cur_index_lso = tx_ring_p->wr_index;
371 	lso_tail_wrap = tx_ring_p->wr_index_wrap;
372 start_again:
373 	ngathers = 0;
374 	sop_index = tx_ring_p->wr_index;
375 #ifdef	NXGE_DEBUG
376 	if (tx_ring_p->descs_pending) {
377 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
378 		    "desc pending %d ", tx_ring_p->descs_pending));
379 	}
380 
381 	dump_len = (int)(MBLKL(mp));
382 	dump_len = (dump_len > 128) ? 128: dump_len;
383 
384 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
385 	    "==> nxge_start: tdc %d: dumping ...: b_rptr $%p "
386 	    "(Before header reserve: ORIGINAL LEN %d)",
387 	    tx_ring_p->tdc,
388 	    mp->b_rptr,
389 	    dump_len));
390 
391 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: dump packets "
392 	    "(IP ORIGINAL b_rptr $%p): %s", mp->b_rptr,
393 	    nxge_dump_packet((char *)mp->b_rptr, dump_len)));
394 #endif
395 
396 	tdc_stats = tx_ring_p->tdc_stats;
397 	mark_mode = (tx_ring_p->descs_pending &&
398 	    ((tx_ring_p->tx_ring_size - tx_ring_p->descs_pending)
399 	    < nxge_tx_minfree));
400 
401 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
402 	    "TX Descriptor ring is channel %d mark mode %d",
403 	    tx_ring_p->tdc, mark_mode));
404 
405 	if ((tx_ring_p->descs_pending + lso_ngathers) >= nxge_reclaim_pending) {
406 		if (!nxge_txdma_reclaim(nxgep, tx_ring_p,
407 		    (nxge_tx_minfree + lso_ngathers))) {
408 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
409 			    "TX Descriptor ring is full: channel %d",
410 			    tx_ring_p->tdc));
411 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
412 			    "TX Descriptor ring is full: channel %d",
413 			    tx_ring_p->tdc));
414 			if (is_lso) {
415 				/*
416 				 * free the current mp and mp_chain if not FULL.
417 				 */
418 				tdc_stats->tx_no_desc++;
419 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
420 				    "LSO packet: TX Descriptor ring is full: "
421 				    "channel %d",
422 				    tx_ring_p->tdc));
423 				goto nxge_start_fail_lso;
424 			} else {
425 				cas32((uint32_t *)&tx_ring_p->queueing, 0, 1);
426 				tdc_stats->tx_no_desc++;
427 
428 				if (isLDOMservice(nxgep)) {
429 					tx_ring_p->tx_ring_busy = B_FALSE;
430 					if (tx_ring_p->tx_ring_offline) {
431 						(void) atomic_swap_32(
432 						    &tx_ring_p->tx_ring_offline,
433 						    NXGE_TX_RING_OFFLINED);
434 					}
435 				}
436 
437 				MUTEX_EXIT(&tx_ring_p->lock);
438 				status = 1;
439 				goto nxge_start_fail1;
440 			}
441 		}
442 	}
443 
444 	nmp = mp;
445 	i = sop_index = tx_ring_p->wr_index;
446 	nmblks = 0;
447 	ngathers = 0;
448 	pkt_len = 0;
449 	pack_len = 0;
450 	clen = 0;
451 	last_bidx = -1;
452 	good_packet = B_TRUE;
453 
454 	desc_area = tx_ring_p->tdc_desc;
455 	npi_handle = desc_area.npi_handle;
456 	npi_desc_handle.regh = (nxge_os_acc_handle_t)
457 	    DMA_COMMON_ACC_HANDLE(desc_area);
458 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
459 	tx_desc_dma_handle = (nxge_os_dma_handle_t)
460 	    DMA_COMMON_HANDLE(desc_area);
461 	tx_msg_ring = tx_ring_p->tx_msg_ring;
462 
463 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: wr_index %d i %d",
464 	    sop_index, i));
465 
466 #ifdef	NXGE_DEBUG
467 	msgsize = msgdsize(nmp);
468 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
469 	    "==> nxge_start(1): wr_index %d i %d msgdsize %d",
470 	    sop_index, i, msgsize));
471 #endif
472 	/*
473 	 * The first 16 bytes of the premapped buffer are reserved
474 	 * for header. No padding will be used.
475 	 */
476 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
477 	if (nxge_tx_use_bcopy && (nxgep->niu_type != N2_NIU)) {
478 		bcopy_thresh = (nxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
479 	} else {
480 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
481 	}
482 	while (nmp) {
483 		good_packet = B_TRUE;
484 		b_rptr = nmp->b_rptr;
485 		len = MBLKL(nmp);
486 		if (len <= 0) {
487 			nmp = nmp->b_cont;
488 			continue;
489 		}
490 		nmblks++;
491 
492 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(1): nmblks %d "
493 		    "len %d pkt_len %d pack_len %d",
494 		    nmblks, len, pkt_len, pack_len));
495 		/*
496 		 * Hardware limits the transfer length to 4K for NIU and
497 		 * 4076 (TX_MAX_TRANSFER_LENGTH) for Neptune. But we just
498 		 * use TX_MAX_TRANSFER_LENGTH as the limit for both.
499 		 * If len is longer than the limit, then we break nmp into
500 		 * two chunks: Make the first chunk equal to the limit and
501 		 * the second chunk for the remaining data. If the second
502 		 * chunk is still larger than the limit, then it will be
503 		 * broken into two in the next pass.
504 		 */
505 		if (len > TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE) {
506 			if ((t_mp = dupb(nmp)) != NULL) {
507 				nmp->b_wptr = nmp->b_rptr +
508 				    (TX_MAX_TRANSFER_LENGTH
509 				    - TX_PKT_HEADER_SIZE);
510 				t_mp->b_rptr = nmp->b_wptr;
511 				t_mp->b_cont = nmp->b_cont;
512 				nmp->b_cont = t_mp;
513 				len = MBLKL(nmp);
514 			} else {
515 				if (is_lso) {
516 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
517 					    "LSO packet: dupb failed: "
518 					    "channel %d",
519 					    tx_ring_p->tdc));
520 					mp = nmp;
521 					goto nxge_start_fail_lso;
522 				} else {
523 					good_packet = B_FALSE;
524 					goto nxge_start_fail2;
525 				}
526 			}
527 		}
528 		tx_desc.value = 0;
529 		tx_desc_p = &tx_desc_ring_vp[i];
530 #ifdef	NXGE_DEBUG
531 		tx_desc_pp = &tx_desc_ring_pp[i];
532 #endif
533 		tx_msg_p = &tx_msg_ring[i];
534 #if defined(__i386)
535 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
536 #else
537 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
538 #endif
539 		if (!header_set &&
540 		    ((!nxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
541 		    (len >= bcopy_thresh))) {
542 			header_set = B_TRUE;
543 			bcopy_thresh += TX_PKT_HEADER_SIZE;
544 			boff = 0;
545 			pack_len = 0;
546 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
547 			hdrp = (p_tx_pkt_header_t)kaddr;
548 			clen = pkt_len;
549 			dma_handle = tx_msg_p->buf_dma_handle;
550 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
551 			(void) ddi_dma_sync(dma_handle,
552 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
553 			    DDI_DMA_SYNC_FORDEV);
554 
555 			tx_msg_p->flags.dma_type = USE_BCOPY;
556 			goto nxge_start_control_header_only;
557 		}
558 
559 		pkt_len += len;
560 		pack_len += len;
561 
562 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(3): "
563 		    "desc entry %d "
564 		    "DESC IOADDR $%p "
565 		    "desc_vp $%p tx_desc_p $%p "
566 		    "desc_pp $%p tx_desc_pp $%p "
567 		    "len %d pkt_len %d pack_len %d",
568 		    i,
569 		    DMA_COMMON_IOADDR(desc_area),
570 		    tx_desc_ring_vp, tx_desc_p,
571 		    tx_desc_ring_pp, tx_desc_pp,
572 		    len, pkt_len, pack_len));
573 
574 		if (len < bcopy_thresh) {
575 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(4): "
576 			    "USE BCOPY: "));
577 			if (nxge_tx_tiny_pack) {
578 				uint32_t blst =
579 				    TXDMA_DESC_NEXT_INDEX(i, -1,
580 				    tx_ring_p->tx_wrap_mask);
581 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
582 				    "==> nxge_start(5): pack"));
583 				if ((pack_len <= bcopy_thresh) &&
584 				    (last_bidx == blst)) {
585 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
586 					    "==> nxge_start: pack(6) "
587 					    "(pkt_len %d pack_len %d)",
588 					    pkt_len, pack_len));
589 					i = blst;
590 					tx_desc_p = &tx_desc_ring_vp[i];
591 #ifdef	NXGE_DEBUG
592 					tx_desc_pp = &tx_desc_ring_pp[i];
593 #endif
594 					tx_msg_p = &tx_msg_ring[i];
595 					boff = pack_len - len;
596 					ngathers--;
597 				} else if (pack_len > bcopy_thresh &&
598 				    header_set) {
599 					pack_len = len;
600 					boff = 0;
601 					bcopy_thresh = nxge_bcopy_thresh;
602 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
603 					    "==> nxge_start(7): > max NEW "
604 					    "bcopy thresh %d "
605 					    "pkt_len %d pack_len %d(next)",
606 					    bcopy_thresh,
607 					    pkt_len, pack_len));
608 				}
609 				last_bidx = i;
610 			}
611 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
612 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
613 				hdrp = (p_tx_pkt_header_t)kaddr;
614 				header_set = B_TRUE;
615 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
616 				    "==> nxge_start(7_x2): "
617 				    "pkt_len %d pack_len %d (new hdrp $%p)",
618 				    pkt_len, pack_len, hdrp));
619 			}
620 			tx_msg_p->flags.dma_type = USE_BCOPY;
621 			kaddr += boff;
622 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(8): "
623 			    "USE BCOPY: before bcopy "
624 			    "DESC IOADDR $%p entry %d "
625 			    "bcopy packets %d "
626 			    "bcopy kaddr $%p "
627 			    "bcopy ioaddr (SAD) $%p "
628 			    "bcopy clen %d "
629 			    "bcopy boff %d",
630 			    DMA_COMMON_IOADDR(desc_area), i,
631 			    tdc_stats->tx_hdr_pkts,
632 			    kaddr,
633 			    dma_ioaddr,
634 			    clen,
635 			    boff));
636 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
637 			    "1USE BCOPY: "));
638 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
639 			    "2USE BCOPY: "));
640 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
641 			    "last USE BCOPY: copy from b_rptr $%p "
642 			    "to KADDR $%p (len %d offset %d",
643 			    b_rptr, kaddr, len, boff));
644 
645 			bcopy(b_rptr, kaddr, len);
646 
647 #ifdef	NXGE_DEBUG
648 			dump_len = (len > 128) ? 128: len;
649 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
650 			    "==> nxge_start: dump packets "
651 			    "(After BCOPY len %d)"
652 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
653 			    nxge_dump_packet((char *)nmp->b_rptr,
654 			    dump_len)));
655 #endif
656 
657 			dma_handle = tx_msg_p->buf_dma_handle;
658 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
659 			(void) ddi_dma_sync(dma_handle,
660 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
661 			    DDI_DMA_SYNC_FORDEV);
662 			clen = len + boff;
663 			tdc_stats->tx_hdr_pkts++;
664 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(9): "
665 			    "USE BCOPY: "
666 			    "DESC IOADDR $%p entry %d "
667 			    "bcopy packets %d "
668 			    "bcopy kaddr $%p "
669 			    "bcopy ioaddr (SAD) $%p "
670 			    "bcopy clen %d "
671 			    "bcopy boff %d",
672 			    DMA_COMMON_IOADDR(desc_area),
673 			    i,
674 			    tdc_stats->tx_hdr_pkts,
675 			    kaddr,
676 			    dma_ioaddr,
677 			    clen,
678 			    boff));
679 		} else {
680 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(12): "
681 			    "USE DVMA: len %d", len));
682 			tx_msg_p->flags.dma_type = USE_DMA;
683 			dma_flags = DDI_DMA_WRITE;
684 			if (len < nxge_dma_stream_thresh) {
685 				dma_flags |= DDI_DMA_CONSISTENT;
686 			} else {
687 				dma_flags |= DDI_DMA_STREAMING;
688 			}
689 
690 			dma_handle = tx_msg_p->dma_handle;
691 			status = ddi_dma_addr_bind_handle(dma_handle, NULL,
692 			    (caddr_t)b_rptr, len, dma_flags,
693 			    DDI_DMA_DONTWAIT, NULL,
694 			    &dma_cookie, &ncookies);
695 			if (status == DDI_DMA_MAPPED) {
696 				dma_ioaddr = dma_cookie.dmac_laddress;
697 				len = (int)dma_cookie.dmac_size;
698 				clen = (uint32_t)dma_cookie.dmac_size;
699 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
700 				    "==> nxge_start(12_1): "
701 				    "USE DVMA: len %d clen %d "
702 				    "ngathers %d",
703 				    len, clen,
704 				    ngathers));
705 #if defined(__i386)
706 				npi_desc_handle.regp = (uint32_t)tx_desc_p;
707 #else
708 				npi_desc_handle.regp = (uint64_t)tx_desc_p;
709 #endif
710 				while (ncookies > 1) {
711 					ngathers++;
712 					/*
713 					 * this is the fix for multiple
714 					 * cookies, which are basically
715 					 * a descriptor entry, we don't set
716 					 * SOP bit as well as related fields
717 					 */
718 
719 					(void) npi_txdma_desc_gather_set(
720 					    npi_desc_handle,
721 					    &tx_desc,
722 					    (ngathers -1),
723 					    mark_mode,
724 					    ngathers,
725 					    dma_ioaddr,
726 					    clen);
727 
728 					tx_msg_p->tx_msg_size = clen;
729 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
730 					    "==> nxge_start:  DMA "
731 					    "ncookie %d "
732 					    "ngathers %d "
733 					    "dma_ioaddr $%p len %d"
734 					    "desc $%p descp $%p (%d)",
735 					    ncookies,
736 					    ngathers,
737 					    dma_ioaddr, clen,
738 					    *tx_desc_p, tx_desc_p, i));
739 
740 					ddi_dma_nextcookie(dma_handle,
741 					    &dma_cookie);
742 					dma_ioaddr =
743 					    dma_cookie.dmac_laddress;
744 
745 					len = (int)dma_cookie.dmac_size;
746 					clen = (uint32_t)dma_cookie.dmac_size;
747 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
748 					    "==> nxge_start(12_2): "
749 					    "USE DVMA: len %d clen %d ",
750 					    len, clen));
751 
752 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
753 					    tx_ring_p->tx_wrap_mask);
754 					tx_desc_p = &tx_desc_ring_vp[i];
755 
756 #if defined(__i386)
757 					npi_desc_handle.regp =
758 					    (uint32_t)tx_desc_p;
759 #else
760 					npi_desc_handle.regp =
761 					    (uint64_t)tx_desc_p;
762 #endif
763 					tx_msg_p = &tx_msg_ring[i];
764 					tx_msg_p->flags.dma_type = USE_NONE;
765 					tx_desc.value = 0;
766 
767 					ncookies--;
768 				}
769 				tdc_stats->tx_ddi_pkts++;
770 				NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start:"
771 				    "DMA: ddi packets %d",
772 				    tdc_stats->tx_ddi_pkts));
773 			} else {
774 				NXGE_ERROR_MSG((nxgep, NXGE_ERR_CTL,
775 				    "dma mapping failed for %d "
776 				    "bytes addr $%p flags %x (%d)",
777 				    len, b_rptr, status, status));
778 				good_packet = B_FALSE;
779 				tdc_stats->tx_dma_bind_fail++;
780 				tx_msg_p->flags.dma_type = USE_NONE;
781 				if (is_lso) {
782 					mp = nmp;
783 					goto nxge_start_fail_lso;
784 				} else {
785 					goto nxge_start_fail2;
786 				}
787 			}
788 		} /* ddi dvma */
789 
790 		if (is_lso) {
791 			nmp_lso_save = nmp;
792 		}
793 		nmp = nmp->b_cont;
794 nxge_start_control_header_only:
795 #if defined(__i386)
796 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
797 #else
798 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
799 #endif
800 		ngathers++;
801 
802 		if (ngathers == 1) {
803 #ifdef	NXGE_DEBUG
804 			save_desc_p = &sop_tx_desc;
805 #endif
806 			sop_tx_desc_p = &sop_tx_desc;
807 			sop_tx_desc_p->value = 0;
808 			sop_tx_desc_p->bits.hdw.tr_len = clen;
809 			sop_tx_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
810 			sop_tx_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
811 		} else {
812 #ifdef	NXGE_DEBUG
813 			save_desc_p = &tx_desc;
814 #endif
815 			tmp_desc_p = &tx_desc;
816 			tmp_desc_p->value = 0;
817 			tmp_desc_p->bits.hdw.tr_len = clen;
818 			tmp_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
819 			tmp_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
820 
821 			tx_desc_p->value = tmp_desc_p->value;
822 		}
823 
824 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(13): "
825 		    "Desc_entry %d ngathers %d "
826 		    "desc_vp $%p tx_desc_p $%p "
827 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
828 		    "dma_ioaddr (SAD) $%p mark %d",
829 		    i, ngathers,
830 		    tx_desc_ring_vp, tx_desc_p,
831 		    len, clen, pkt_len, pack_len, nmblks,
832 		    dma_ioaddr, mark_mode));
833 
834 #ifdef NXGE_DEBUG
835 		npi_desc_handle.nxgep = nxgep;
836 		npi_desc_handle.function.function = nxgep->function_num;
837 		npi_desc_handle.function.instance = nxgep->instance;
838 		sad = (save_desc_p->value & TX_PKT_DESC_SAD_MASK);
839 		xfer_len = ((save_desc_p->value & TX_PKT_DESC_TR_LEN_MASK) >>
840 		    TX_PKT_DESC_TR_LEN_SHIFT);
841 
842 
843 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
844 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
845 		    "mark %d sop %d\n",
846 		    save_desc_p->value,
847 		    sad,
848 		    save_desc_p->bits.hdw.tr_len,
849 		    xfer_len,
850 		    save_desc_p->bits.hdw.num_ptr,
851 		    save_desc_p->bits.hdw.mark,
852 		    save_desc_p->bits.hdw.sop));
853 
854 		npi_txdma_dump_desc_one(npi_desc_handle, NULL, i);
855 #endif
856 
857 		tx_msg_p->tx_msg_size = clen;
858 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
859 		if (ngathers > nxge_tx_max_gathers) {
860 			good_packet = B_FALSE;
861 			hcksum_retrieve(mp, NULL, NULL, &start_offset,
862 			    &stuff_offset, &end_offset, &value,
863 			    &cksum_flags);
864 
865 			NXGE_DEBUG_MSG((NULL, TX_CTL,
866 			    "==> nxge_start(14): pull msg - "
867 			    "len %d pkt_len %d ngathers %d",
868 			    len, pkt_len, ngathers));
869 			/* Pull all message blocks from b_cont */
870 			if (is_lso) {
871 				mp = nmp_lso_save;
872 				goto nxge_start_fail_lso;
873 			}
874 			if ((msgpullup(mp, -1)) == NULL) {
875 				goto nxge_start_fail2;
876 			}
877 			goto nxge_start_fail2;
878 		}
879 	} /* while (nmp) */
880 
881 	tx_msg_p->tx_message = mp;
882 	tx_desc_p = &tx_desc_ring_vp[sop_index];
883 #if defined(__i386)
884 	npi_desc_handle.regp = (uint32_t)tx_desc_p;
885 #else
886 	npi_desc_handle.regp = (uint64_t)tx_desc_p;
887 #endif
888 
889 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
890 	pkthdrp->reserved = 0;
891 	hdrp->value = 0;
892 	bcopy(&tmp_hdrp, hdrp, sizeof (tx_pkt_header_t));
893 
894 	if (pkt_len > NXGE_MTU_DEFAULT_MAX) {
895 		tdc_stats->tx_jumbo_pkts++;
896 	}
897 
898 	min_len = (ETHERMIN + TX_PKT_HEADER_SIZE + (npads * 2));
899 	if (pkt_len < min_len) {
900 		/* Assume we use bcopy to premapped buffers */
901 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
902 		NXGE_DEBUG_MSG((NULL, TX_CTL,
903 		    "==> nxge_start(14-1): < (msg_min + 16)"
904 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
905 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
906 		bzero((kaddr + pkt_len), (min_len - pkt_len));
907 		pkt_len = tx_msg_p->tx_msg_size = min_len;
908 
909 		sop_tx_desc_p->bits.hdw.tr_len = min_len;
910 
911 		NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
912 		tx_desc_p->value = sop_tx_desc_p->value;
913 
914 		NXGE_DEBUG_MSG((NULL, TX_CTL,
915 		    "==> nxge_start(14-2): < msg_min - "
916 		    "len %d pkt_len %d min_len %d ngathers %d",
917 		    len, pkt_len, min_len, ngathers));
918 	}
919 
920 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: cksum_flags 0x%x ",
921 	    cksum_flags));
922 	{
923 		uint64_t	tmp_len;
924 
925 		/* pkt_len already includes 16 + paddings!! */
926 		/* Update the control header length */
927 		tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
928 		tmp_len = hdrp->value |
929 		    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
930 
931 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
932 		    "==> nxge_start(15_x1): setting SOP "
933 		    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
934 		    "0x%llx hdrp->value 0x%llx",
935 		    tot_xfer_len, tot_xfer_len, pkt_len,
936 		    tmp_len, hdrp->value));
937 #if defined(_BIG_ENDIAN)
938 		hdrp->value = ddi_swap64(tmp_len);
939 #else
940 		hdrp->value = tmp_len;
941 #endif
942 		NXGE_DEBUG_MSG((nxgep,
943 		    TX_CTL, "==> nxge_start(15_x2): setting SOP "
944 		    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
945 		    "tmp_len 0x%llx hdrp->value 0x%llx",
946 		    tot_xfer_len, pkt_len,
947 		    tmp_len, hdrp->value));
948 	}
949 
950 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(15): setting SOP "
951 	    "wr_index %d "
952 	    "tot_xfer_len (%d) pkt_len %d npads %d",
953 	    sop_index,
954 	    tot_xfer_len, pkt_len,
955 	    npads));
956 
957 	sop_tx_desc_p->bits.hdw.sop = 1;
958 	sop_tx_desc_p->bits.hdw.mark = mark_mode;
959 	sop_tx_desc_p->bits.hdw.num_ptr = ngathers;
960 
961 	NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
962 
963 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(16): set SOP done"));
964 
965 #ifdef NXGE_DEBUG
966 	npi_desc_handle.nxgep = nxgep;
967 	npi_desc_handle.function.function = nxgep->function_num;
968 	npi_desc_handle.function.instance = nxgep->instance;
969 
970 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
971 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
972 	    save_desc_p->value,
973 	    sad,
974 	    save_desc_p->bits.hdw.tr_len,
975 	    xfer_len,
976 	    save_desc_p->bits.hdw.num_ptr,
977 	    save_desc_p->bits.hdw.mark,
978 	    save_desc_p->bits.hdw.sop));
979 	(void) npi_txdma_dump_desc_one(npi_desc_handle, NULL, sop_index);
980 
981 	dump_len = (pkt_len > 128) ? 128: pkt_len;
982 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
983 	    "==> nxge_start: dump packets(17) (after sop set, len "
984 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
985 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
986 	    (char *)hdrp,
987 	    nxge_dump_packet((char *)hdrp, dump_len)));
988 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
989 	    "==> nxge_start(18): TX desc sync: sop_index %d",
990 	    sop_index));
991 #endif
992 
993 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
994 		(void) ddi_dma_sync(tx_desc_dma_handle,
995 		    sop_index * sizeof (tx_desc_t),
996 		    ngathers * sizeof (tx_desc_t),
997 		    DDI_DMA_SYNC_FORDEV);
998 
999 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(19): sync 1 "
1000 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1001 		    "pkt_len %d ngathers %d sop_index %d\n",
1002 		    stuff_offset, start_offset,
1003 		    pkt_len, ngathers, sop_index));
1004 	} else { /* more than one descriptor and wrap around */
1005 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
1006 		(void) ddi_dma_sync(tx_desc_dma_handle,
1007 		    sop_index * sizeof (tx_desc_t),
1008 		    nsdescs * sizeof (tx_desc_t),
1009 		    DDI_DMA_SYNC_FORDEV);
1010 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(20): sync 1 "
1011 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1012 		    "pkt_len %d ngathers %d sop_index %d\n",
1013 		    stuff_offset, start_offset,
1014 		    pkt_len, ngathers, sop_index));
1015 
1016 		(void) ddi_dma_sync(tx_desc_dma_handle,
1017 		    0,
1018 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
1019 		    DDI_DMA_SYNC_FORDEV);
1020 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(21): sync 2 "
1021 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1022 		    "pkt_len %d ngathers %d sop_index %d\n",
1023 		    stuff_offset, start_offset,
1024 		    pkt_len, ngathers, sop_index));
1025 	}
1026 
1027 	tail_index = tx_ring_p->wr_index;
1028 	tail_wrap = tx_ring_p->wr_index_wrap;
1029 
1030 	tx_ring_p->wr_index = i;
1031 	if (tx_ring_p->wr_index <= tail_index) {
1032 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
1033 		    B_FALSE : B_TRUE);
1034 	}
1035 
1036 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX kick: "
1037 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
1038 	    tx_ring_p->tdc,
1039 	    tx_ring_p->wr_index,
1040 	    tx_ring_p->wr_index_wrap,
1041 	    ngathers,
1042 	    tx_ring_p->descs_pending));
1043 
1044 	if (is_lso) {
1045 		lso_ngathers += ngathers;
1046 		if (mp_chain != NULL) {
1047 			mp = mp_chain;
1048 			mp_chain = mp_chain->b_next;
1049 			mp->b_next = NULL;
1050 			if (nxge_lso_kick_cnt == lso_ngathers) {
1051 				tx_ring_p->descs_pending += lso_ngathers;
1052 				{
1053 					tx_ring_kick_t		kick;
1054 
1055 					kick.value = 0;
1056 					kick.bits.ldw.wrap =
1057 					    tx_ring_p->wr_index_wrap;
1058 					kick.bits.ldw.tail =
1059 					    (uint16_t)tx_ring_p->wr_index;
1060 
1061 					/* Kick the Transmit kick register */
1062 					TXDMA_REG_WRITE64(
1063 					    NXGE_DEV_NPI_HANDLE(nxgep),
1064 					    TX_RING_KICK_REG,
1065 					    (uint8_t)tx_ring_p->tdc,
1066 					    kick.value);
1067 					tdc_stats->tx_starts++;
1068 
1069 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
1070 					    "==> nxge_start: more LSO: "
1071 					    "LSO_CNT %d",
1072 					    lso_ngathers));
1073 				}
1074 				lso_ngathers = 0;
1075 				ngathers = 0;
1076 				cur_index_lso = sop_index = tx_ring_p->wr_index;
1077 				lso_tail_wrap = tx_ring_p->wr_index_wrap;
1078 			}
1079 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1080 			    "==> nxge_start: lso again: "
1081 			    "lso_gathers %d ngathers %d cur_index_lso %d "
1082 			    "wr_index %d sop_index %d",
1083 			    lso_ngathers, ngathers, cur_index_lso,
1084 			    tx_ring_p->wr_index, sop_index));
1085 
1086 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1087 			    "==> nxge_start: next : count %d",
1088 			    lso_ngathers));
1089 			lso_again = B_TRUE;
1090 			goto start_again;
1091 		}
1092 		ngathers = lso_ngathers;
1093 	}
1094 
1095 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX KICKING: "));
1096 
1097 	{
1098 		tx_ring_kick_t		kick;
1099 
1100 		kick.value = 0;
1101 		kick.bits.ldw.wrap = tx_ring_p->wr_index_wrap;
1102 		kick.bits.ldw.tail = (uint16_t)tx_ring_p->wr_index;
1103 
1104 		/* Kick start the Transmit kick register */
1105 		TXDMA_REG_WRITE64(NXGE_DEV_NPI_HANDLE(nxgep),
1106 		    TX_RING_KICK_REG,
1107 		    (uint8_t)tx_ring_p->tdc,
1108 		    kick.value);
1109 	}
1110 
1111 	tx_ring_p->descs_pending += ngathers;
1112 	tdc_stats->tx_starts++;
1113 
1114 	if (isLDOMservice(nxgep)) {
1115 		tx_ring_p->tx_ring_busy = B_FALSE;
1116 		if (tx_ring_p->tx_ring_offline) {
1117 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1118 			    NXGE_TX_RING_OFFLINED);
1119 		}
1120 	}
1121 
1122 	MUTEX_EXIT(&tx_ring_p->lock);
1123 
1124 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1125 	return (status);
1126 
1127 nxge_start_fail_lso:
1128 	status = 0;
1129 	good_packet = B_FALSE;
1130 	if (mp != NULL) {
1131 		freemsg(mp);
1132 	}
1133 	if (mp_chain != NULL) {
1134 		freemsg(mp_chain);
1135 	}
1136 	if (!lso_again && !ngathers) {
1137 		if (isLDOMservice(nxgep)) {
1138 			tx_ring_p->tx_ring_busy = B_FALSE;
1139 			if (tx_ring_p->tx_ring_offline) {
1140 				(void) atomic_swap_32(
1141 				    &tx_ring_p->tx_ring_offline,
1142 				    NXGE_TX_RING_OFFLINED);
1143 			}
1144 		}
1145 
1146 		MUTEX_EXIT(&tx_ring_p->lock);
1147 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
1148 		    "==> nxge_start: lso exit (nothing changed)"));
1149 		goto nxge_start_fail1;
1150 	}
1151 
1152 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1153 	    "==> nxge_start (channel %d): before lso "
1154 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1155 	    "wr_index %d sop_index %d lso_again %d",
1156 	    tx_ring_p->tdc,
1157 	    lso_ngathers, ngathers, cur_index_lso,
1158 	    tx_ring_p->wr_index, sop_index, lso_again));
1159 
1160 	if (lso_again) {
1161 		lso_ngathers += ngathers;
1162 		ngathers = lso_ngathers;
1163 		sop_index = cur_index_lso;
1164 		tx_ring_p->wr_index = sop_index;
1165 		tx_ring_p->wr_index_wrap = lso_tail_wrap;
1166 	}
1167 
1168 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1169 	    "==> nxge_start (channel %d): after lso "
1170 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1171 	    "wr_index %d sop_index %d lso_again %d",
1172 	    tx_ring_p->tdc,
1173 	    lso_ngathers, ngathers, cur_index_lso,
1174 	    tx_ring_p->wr_index, sop_index, lso_again));
1175 
1176 nxge_start_fail2:
1177 	if (good_packet == B_FALSE) {
1178 		cur_index = sop_index;
1179 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: clean up"));
1180 		for (i = 0; i < ngathers; i++) {
1181 			tx_desc_p = &tx_desc_ring_vp[cur_index];
1182 #if defined(__i386)
1183 			npi_handle.regp = (uint32_t)tx_desc_p;
1184 #else
1185 			npi_handle.regp = (uint64_t)tx_desc_p;
1186 #endif
1187 			tx_msg_p = &tx_msg_ring[cur_index];
1188 			(void) npi_txdma_desc_set_zero(npi_handle, 1);
1189 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
1190 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
1191 				    "tx_desc_p = %X index = %d",
1192 				    tx_desc_p, tx_ring_p->rd_index));
1193 				(void) dvma_unload(tx_msg_p->dvma_handle,
1194 				    0, -1);
1195 				tx_msg_p->dvma_handle = NULL;
1196 				if (tx_ring_p->dvma_wr_index ==
1197 				    tx_ring_p->dvma_wrap_mask)
1198 					tx_ring_p->dvma_wr_index = 0;
1199 				else
1200 					tx_ring_p->dvma_wr_index++;
1201 				tx_ring_p->dvma_pending--;
1202 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
1203 				if (ddi_dma_unbind_handle(
1204 				    tx_msg_p->dma_handle)) {
1205 					cmn_err(CE_WARN, "!nxge_start: "
1206 					    "ddi_dma_unbind_handle failed");
1207 				}
1208 			}
1209 			tx_msg_p->flags.dma_type = USE_NONE;
1210 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
1211 			    tx_ring_p->tx_wrap_mask);
1212 
1213 		}
1214 	}
1215 
1216 	if (isLDOMservice(nxgep)) {
1217 		tx_ring_p->tx_ring_busy = B_FALSE;
1218 		if (tx_ring_p->tx_ring_offline) {
1219 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1220 			    NXGE_TX_RING_OFFLINED);
1221 		}
1222 	}
1223 
1224 	MUTEX_EXIT(&tx_ring_p->lock);
1225 
1226 nxge_start_fail1:
1227 	/* Add FMA to check the access handle nxge_hregh */
1228 
1229 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1230 	return (status);
1231 }
1232 
1233 /* Software LSO starts here */
1234 static void
1235 nxge_hcksum_retrieve(mblk_t *mp,
1236     uint32_t *start, uint32_t *stuff, uint32_t *end,
1237     uint32_t *value, uint32_t *flags)
1238 {
1239 	if (mp->b_datap->db_type == M_DATA) {
1240 		if (flags != NULL) {
1241 			*flags = DB_CKSUMFLAGS(mp) & (HCK_IPV4_HDRCKSUM |
1242 			    HCK_PARTIALCKSUM | HCK_FULLCKSUM |
1243 			    HCK_FULLCKSUM_OK);
1244 			if ((*flags & (HCK_PARTIALCKSUM |
1245 			    HCK_FULLCKSUM)) != 0) {
1246 				if (value != NULL)
1247 					*value = (uint32_t)DB_CKSUM16(mp);
1248 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
1249 					if (start != NULL)
1250 						*start =
1251 						    (uint32_t)DB_CKSUMSTART(mp);
1252 					if (stuff != NULL)
1253 						*stuff =
1254 						    (uint32_t)DB_CKSUMSTUFF(mp);
1255 					if (end != NULL)
1256 						*end =
1257 						    (uint32_t)DB_CKSUMEND(mp);
1258 				}
1259 			}
1260 		}
1261 	}
1262 }
1263 
1264 static void
1265 nxge_lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
1266 {
1267 	ASSERT(DB_TYPE(mp) == M_DATA);
1268 
1269 	*mss = 0;
1270 	if (flags != NULL) {
1271 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO;
1272 		if ((*flags != 0) && (mss != NULL)) {
1273 			*mss = (uint32_t)DB_LSOMSS(mp);
1274 		}
1275 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1276 		    "==> nxge_lso_info_get(flag !=NULL): mss %d *flags 0x%x",
1277 		    *mss, *flags));
1278 	}
1279 
1280 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1281 	    "<== nxge_lso_info_get: mss %d", *mss));
1282 }
1283 
1284 /*
1285  * Do Soft LSO on the oversized packet.
1286  *
1287  * 1. Create a chain of message for headers.
1288  * 2. Fill up header messages with proper information.
1289  * 3. Copy Eithernet, IP, and TCP headers from the original message to
1290  *    each new message with necessary adjustments.
1291  *    * Unchange the ethernet header for DIX frames. (by default)
1292  *    * IP Total Length field is updated to MSS or less(only for the last one).
1293  *    * IP Identification value is incremented by one for each packet.
1294  *    * TCP sequence Number is recalculated according to the payload length.
1295  *    * Set FIN and/or PSH flags for the *last* packet if applied.
1296  *    * TCP partial Checksum
1297  * 4. Update LSO information in the first message header.
1298  * 5. Release the original message header.
1299  */
1300 static mblk_t *
1301 nxge_do_softlso(mblk_t *mp, uint32_t mss)
1302 {
1303 	uint32_t	hckflags;
1304 	int		pktlen;
1305 	int		hdrlen;
1306 	int		segnum;
1307 	int		i;
1308 	struct ether_vlan_header *evh;
1309 	int		ehlen, iphlen, tcphlen;
1310 	struct ip	*oiph, *niph;
1311 	struct tcphdr *otcph, *ntcph;
1312 	int		available, len, left;
1313 	uint16_t	ip_id;
1314 	uint32_t	tcp_seq;
1315 #ifdef __sparc
1316 	uint32_t	tcp_seq_tmp;
1317 #endif
1318 	mblk_t		*datamp;
1319 	uchar_t		*rptr;
1320 	mblk_t		*nmp;
1321 	mblk_t		*cmp;
1322 	mblk_t		*mp_chain;
1323 	boolean_t do_cleanup = B_FALSE;
1324 	t_uscalar_t start_offset = 0;
1325 	t_uscalar_t stuff_offset = 0;
1326 	t_uscalar_t value = 0;
1327 	uint16_t	l4_len;
1328 	ipaddr_t	src, dst;
1329 	uint32_t	cksum, sum, l4cksum;
1330 
1331 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1332 	    "==> nxge_do_softlso"));
1333 	/*
1334 	 * check the length of LSO packet payload and calculate the number of
1335 	 * segments to be generated.
1336 	 */
1337 	pktlen = msgsize(mp);
1338 	evh = (struct ether_vlan_header *)mp->b_rptr;
1339 
1340 	/* VLAN? */
1341 	if (evh->ether_tpid == htons(ETHERTYPE_VLAN))
1342 		ehlen = sizeof (struct ether_vlan_header);
1343 	else
1344 		ehlen = sizeof (struct ether_header);
1345 	oiph = (struct ip *)(mp->b_rptr + ehlen);
1346 	iphlen = oiph->ip_hl * 4;
1347 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1348 	tcphlen = otcph->th_off * 4;
1349 
1350 	l4_len = pktlen - ehlen - iphlen;
1351 
1352 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1353 	    "==> nxge_do_softlso: mss %d oiph $%p "
1354 	    "original ip_sum oiph->ip_sum 0x%x "
1355 	    "original tcp_sum otcph->th_sum 0x%x "
1356 	    "oiph->ip_len %d pktlen %d ehlen %d "
1357 	    "l4_len %d (0x%x) ip_len - iphlen %d ",
1358 	    mss,
1359 	    oiph,
1360 	    oiph->ip_sum,
1361 	    otcph->th_sum,
1362 	    ntohs(oiph->ip_len), pktlen,
1363 	    ehlen,
1364 	    l4_len,
1365 	    l4_len,
1366 	    ntohs(oiph->ip_len) - iphlen));
1367 
1368 	/* IPv4 + TCP */
1369 	if (!(oiph->ip_v == IPV4_VERSION)) {
1370 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1371 		    "<== nxge_do_softlso: not IPV4 "
1372 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1373 		    ntohs(oiph->ip_len), pktlen, ehlen,
1374 		    tcphlen));
1375 		freemsg(mp);
1376 		return (NULL);
1377 	}
1378 
1379 	if (!(oiph->ip_p == IPPROTO_TCP)) {
1380 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1381 		    "<== nxge_do_softlso: not TCP "
1382 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1383 		    ntohs(oiph->ip_len), pktlen, ehlen,
1384 		    tcphlen));
1385 		freemsg(mp);
1386 		return (NULL);
1387 	}
1388 
1389 	if (!(ntohs(oiph->ip_len) == pktlen - ehlen)) {
1390 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1391 		    "<== nxge_do_softlso: len not matched  "
1392 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1393 		    ntohs(oiph->ip_len), pktlen, ehlen,
1394 		    tcphlen));
1395 		freemsg(mp);
1396 		return (NULL);
1397 	}
1398 
1399 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1400 	tcphlen = otcph->th_off * 4;
1401 
1402 	/* TCP flags can not include URG, RST, or SYN */
1403 	VERIFY((otcph->th_flags & (TH_SYN | TH_RST | TH_URG)) == 0);
1404 
1405 	hdrlen = ehlen + iphlen + tcphlen;
1406 
1407 	VERIFY(MBLKL(mp) >= hdrlen);
1408 
1409 	if (MBLKL(mp) > hdrlen) {
1410 		datamp = mp;
1411 		rptr = mp->b_rptr + hdrlen;
1412 	} else { /* = */
1413 		datamp = mp->b_cont;
1414 		rptr = datamp->b_rptr;
1415 	}
1416 
1417 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1418 	    "nxge_do_softlso: otcph $%p pktlen: %d, "
1419 	    "hdrlen %d ehlen %d iphlen %d tcphlen %d "
1420 	    "mblkl(mp): %d, mblkl(datamp): %d",
1421 	    otcph,
1422 	    pktlen, hdrlen, ehlen, iphlen, tcphlen,
1423 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1424 
1425 	hckflags = 0;
1426 	nxge_hcksum_retrieve(mp,
1427 	    &start_offset, &stuff_offset, &value, NULL, &hckflags);
1428 
1429 	dst = oiph->ip_dst.s_addr;
1430 	src = oiph->ip_src.s_addr;
1431 
1432 	cksum = (dst >> 16) + (dst & 0xFFFF) +
1433 	    (src >> 16) + (src & 0xFFFF);
1434 	l4cksum = cksum + IP_TCP_CSUM_COMP;
1435 
1436 	sum = l4_len + l4cksum;
1437 	sum = (sum & 0xFFFF) + (sum >> 16);
1438 
1439 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1440 	    "==> nxge_do_softlso: dst 0x%x src 0x%x sum 0x%x ~new 0x%x "
1441 	    "hckflags 0x%x start_offset %d stuff_offset %d "
1442 	    "value (original) 0x%x th_sum 0x%x "
1443 	    "pktlen %d l4_len %d (0x%x) "
1444 	    "MBLKL(mp): %d, MBLKL(datamp): %d dump header %s",
1445 	    dst, src,
1446 	    (sum & 0xffff), (~sum & 0xffff),
1447 	    hckflags, start_offset, stuff_offset,
1448 	    value, otcph->th_sum,
1449 	    pktlen,
1450 	    l4_len,
1451 	    l4_len,
1452 	    ntohs(oiph->ip_len) - (int)MBLKL(mp),
1453 	    (int)MBLKL(datamp),
1454 	    nxge_dump_packet((char *)evh, 12)));
1455 
1456 	/*
1457 	 * Start to process.
1458 	 */
1459 	available = pktlen - hdrlen;
1460 	segnum = (available - 1) / mss + 1;
1461 
1462 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1463 	    "==> nxge_do_softlso: pktlen %d "
1464 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1465 	    "available %d mss %d segnum %d",
1466 	    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp),
1467 	    available,
1468 	    mss,
1469 	    segnum));
1470 
1471 	VERIFY(segnum >= 2);
1472 
1473 	/*
1474 	 * Try to pre-allocate all header messages
1475 	 */
1476 	mp_chain = NULL;
1477 	for (i = 0; i < segnum; i++) {
1478 		if ((nmp = allocb(hdrlen, 0)) == NULL) {
1479 			/* Clean up the mp_chain */
1480 			while (mp_chain != NULL) {
1481 				nmp = mp_chain;
1482 				mp_chain = mp_chain->b_next;
1483 				freemsg(nmp);
1484 			}
1485 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1486 			    "<== nxge_do_softlso: "
1487 			    "Could not allocate enough messages for headers!"));
1488 			freemsg(mp);
1489 			return (NULL);
1490 		}
1491 		nmp->b_next = mp_chain;
1492 		mp_chain = nmp;
1493 
1494 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1495 		    "==> nxge_do_softlso: "
1496 		    "mp $%p nmp $%p mp_chain $%p mp_chain->b_next $%p",
1497 		    mp, nmp, mp_chain, mp_chain->b_next));
1498 	}
1499 
1500 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1501 	    "==> nxge_do_softlso: mp $%p nmp $%p mp_chain $%p",
1502 	    mp, nmp, mp_chain));
1503 
1504 	/*
1505 	 * Associate payload with new packets
1506 	 */
1507 	cmp = mp_chain;
1508 	left = available;
1509 	while (cmp != NULL) {
1510 		nmp = dupb(datamp);
1511 		if (nmp == NULL) {
1512 			do_cleanup = B_TRUE;
1513 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1514 			    "==>nxge_do_softlso: "
1515 			    "Can not dupb(datamp), have to do clean up"));
1516 			goto cleanup_allocated_msgs;
1517 		}
1518 
1519 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1520 		    "==> nxge_do_softlso: (loop) before mp $%p cmp $%p "
1521 		    "dupb nmp $%p len %d left %d msd %d ",
1522 		    mp, cmp, nmp, len, left, mss));
1523 
1524 		cmp->b_cont = nmp;
1525 		nmp->b_rptr = rptr;
1526 		len = (left < mss) ? left : mss;
1527 		left -= len;
1528 
1529 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1530 		    "==> nxge_do_softlso: (loop) after mp $%p cmp $%p "
1531 		    "dupb nmp $%p len %d left %d mss %d ",
1532 		    mp, cmp, nmp, len, left, mss));
1533 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1534 		    "nxge_do_softlso: before available: %d, "
1535 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1536 		    available, left, len, segnum, (int)MBLKL(nmp)));
1537 
1538 		len -= MBLKL(nmp);
1539 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1540 		    "nxge_do_softlso: after available: %d, "
1541 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1542 		    available, left, len, segnum, (int)MBLKL(nmp)));
1543 
1544 		while (len > 0) {
1545 			mblk_t *mmp = NULL;
1546 
1547 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1548 			    "nxge_do_softlso: (4) len > 0 available: %d, "
1549 			    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1550 			    available, left, len, segnum, (int)MBLKL(nmp)));
1551 
1552 			if (datamp->b_cont != NULL) {
1553 				datamp = datamp->b_cont;
1554 				rptr = datamp->b_rptr;
1555 				mmp = dupb(datamp);
1556 				if (mmp == NULL) {
1557 					do_cleanup = B_TRUE;
1558 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1559 					    "==> nxge_do_softlso: "
1560 					    "Can not dupb(datamp) (1), :"
1561 					    "have to do clean up"));
1562 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1563 					    "==> nxge_do_softlso: "
1564 					    "available: %d, left: %d, "
1565 					    "len: %d, MBLKL(nmp): %d",
1566 					    available, left, len,
1567 					    (int)MBLKL(nmp)));
1568 					goto cleanup_allocated_msgs;
1569 				}
1570 			} else {
1571 				NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1572 				    "==> nxge_do_softlso: "
1573 				    "(1)available: %d, left: %d, "
1574 				    "len: %d, MBLKL(nmp): %d",
1575 				    available, left, len,
1576 				    (int)MBLKL(nmp)));
1577 				cmn_err(CE_PANIC,
1578 				    "==> nxge_do_softlso: "
1579 				    "Pointers must have been corrupted!\n"
1580 				    "datamp: $%p, nmp: $%p, rptr: $%p",
1581 				    (void *)datamp,
1582 				    (void *)nmp,
1583 				    (void *)rptr);
1584 			}
1585 			nmp->b_cont = mmp;
1586 			nmp = mmp;
1587 			len -= MBLKL(nmp);
1588 		}
1589 		if (len < 0) {
1590 			nmp->b_wptr += len;
1591 			rptr = nmp->b_wptr;
1592 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1593 			    "(5) len < 0 (less than 0)"
1594 			    "available: %d, left: %d, len: %d, MBLKL(nmp): %d",
1595 			    available, left, len, (int)MBLKL(nmp)));
1596 
1597 		} else if (len == 0) {
1598 			if (datamp->b_cont != NULL) {
1599 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1600 				    "(5) len == 0"
1601 				    "available: %d, left: %d, len: %d, "
1602 				    "MBLKL(nmp): %d",
1603 				    available, left, len, (int)MBLKL(nmp)));
1604 				datamp = datamp->b_cont;
1605 				rptr = datamp->b_rptr;
1606 			} else {
1607 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1608 				    "(6)available b_cont == NULL : %d, "
1609 				    "left: %d, len: %d, MBLKL(nmp): %d",
1610 				    available, left, len, (int)MBLKL(nmp)));
1611 
1612 				VERIFY(cmp->b_next == NULL);
1613 				VERIFY(left == 0);
1614 				break; /* Done! */
1615 			}
1616 		}
1617 		cmp = cmp->b_next;
1618 
1619 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1620 		    "(7) do_softlso: "
1621 		    "next mp in mp_chain available len != 0 : %d, "
1622 		    "left: %d, len: %d, MBLKL(nmp): %d",
1623 		    available, left, len, (int)MBLKL(nmp)));
1624 	}
1625 
1626 	/*
1627 	 * From now, start to fill up all headers for the first message
1628 	 * Hardware checksum flags need to be updated separately for FULLCKSUM
1629 	 * and PARTIALCKSUM cases. For full checksum, copy the original flags
1630 	 * into every new packet is enough. But for HCK_PARTIALCKSUM, all
1631 	 * required fields need to be updated properly.
1632 	 */
1633 	nmp = mp_chain;
1634 	bcopy(mp->b_rptr, nmp->b_rptr, hdrlen);
1635 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1636 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1637 	niph->ip_len = htons(mss + iphlen + tcphlen);
1638 	ip_id = ntohs(niph->ip_id);
1639 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1640 #ifdef __sparc
1641 	bcopy((char *)&ntcph->th_seq, &tcp_seq_tmp, 4);
1642 	tcp_seq = ntohl(tcp_seq_tmp);
1643 #else
1644 	tcp_seq = ntohl(ntcph->th_seq);
1645 #endif
1646 
1647 	ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST);
1648 
1649 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1650 	DB_CKSUMSTART(nmp) = start_offset;
1651 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1652 
1653 	/* calculate IP checksum and TCP pseudo header checksum */
1654 	niph->ip_sum = 0;
1655 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1656 
1657 	l4_len = mss + tcphlen;
1658 	sum = htons(l4_len) + l4cksum;
1659 	sum = (sum & 0xFFFF) + (sum >> 16);
1660 	ntcph->th_sum = (sum & 0xffff);
1661 
1662 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1663 	    "==> nxge_do_softlso: first mp $%p (mp_chain $%p) "
1664 	    "mss %d pktlen %d l4_len %d (0x%x) "
1665 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1666 	    "ip_sum 0x%x "
1667 	    "th_sum 0x%x sum 0x%x ) "
1668 	    "dump first ip->tcp %s",
1669 	    nmp, mp_chain,
1670 	    mss,
1671 	    pktlen,
1672 	    l4_len,
1673 	    l4_len,
1674 	    (int)MBLKL(mp), (int)MBLKL(datamp),
1675 	    niph->ip_sum,
1676 	    ntcph->th_sum,
1677 	    sum,
1678 	    nxge_dump_packet((char *)niph, 52)));
1679 
1680 	cmp = nmp;
1681 	while ((nmp = nmp->b_next)->b_next != NULL) {
1682 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1683 		    "==>nxge_do_softlso: middle l4_len %d ", l4_len));
1684 		bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1685 		nmp->b_wptr = nmp->b_rptr + hdrlen;
1686 		niph = (struct ip *)(nmp->b_rptr + ehlen);
1687 		niph->ip_id = htons(++ip_id);
1688 		niph->ip_len = htons(mss + iphlen + tcphlen);
1689 		ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1690 		tcp_seq += mss;
1691 
1692 		ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST | TH_URG);
1693 
1694 #ifdef __sparc
1695 		tcp_seq_tmp = htonl(tcp_seq);
1696 		bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1697 #else
1698 		ntcph->th_seq = htonl(tcp_seq);
1699 #endif
1700 		DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1701 		DB_CKSUMSTART(nmp) = start_offset;
1702 		DB_CKSUMSTUFF(nmp) = stuff_offset;
1703 
1704 		/* calculate IP checksum and TCP pseudo header checksum */
1705 		niph->ip_sum = 0;
1706 		niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1707 		ntcph->th_sum = (sum & 0xffff);
1708 
1709 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1710 		    "==> nxge_do_softlso: middle ip_sum 0x%x "
1711 		    "th_sum 0x%x "
1712 		    " mp $%p (mp_chain $%p) pktlen %d "
1713 		    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1714 		    niph->ip_sum,
1715 		    ntcph->th_sum,
1716 		    nmp, mp_chain,
1717 		    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp)));
1718 	}
1719 
1720 	/* Last segment */
1721 	/*
1722 	 * Set FIN and/or PSH flags if present only in the last packet.
1723 	 * The ip_len could be different from prior packets.
1724 	 */
1725 	bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1726 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1727 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1728 	niph->ip_id = htons(++ip_id);
1729 	niph->ip_len = htons(msgsize(nmp->b_cont) + iphlen + tcphlen);
1730 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1731 	tcp_seq += mss;
1732 #ifdef __sparc
1733 	tcp_seq_tmp = htonl(tcp_seq);
1734 	bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1735 #else
1736 	ntcph->th_seq = htonl(tcp_seq);
1737 #endif
1738 	ntcph->th_flags = (otcph->th_flags & ~TH_URG);
1739 
1740 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1741 	DB_CKSUMSTART(nmp) = start_offset;
1742 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1743 
1744 	/* calculate IP checksum and TCP pseudo header checksum */
1745 	niph->ip_sum = 0;
1746 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1747 
1748 	l4_len = ntohs(niph->ip_len) - iphlen;
1749 	sum = htons(l4_len) + l4cksum;
1750 	sum = (sum & 0xFFFF) + (sum >> 16);
1751 	ntcph->th_sum = (sum & 0xffff);
1752 
1753 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1754 	    "==> nxge_do_softlso: last next "
1755 	    "niph->ip_sum 0x%x "
1756 	    "ntcph->th_sum 0x%x sum 0x%x "
1757 	    "dump last ip->tcp %s "
1758 	    "cmp $%p mp $%p (mp_chain $%p) pktlen %d (0x%x) "
1759 	    "l4_len %d (0x%x) "
1760 	    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1761 	    niph->ip_sum,
1762 	    ntcph->th_sum, sum,
1763 	    nxge_dump_packet((char *)niph, 52),
1764 	    cmp, nmp, mp_chain,
1765 	    pktlen, pktlen,
1766 	    l4_len,
1767 	    l4_len,
1768 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1769 
1770 cleanup_allocated_msgs:
1771 	if (do_cleanup) {
1772 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1773 		    "==> nxge_do_softlso: "
1774 		    "Failed allocating messages, "
1775 		    "have to clean up and fail!"));
1776 		while (mp_chain != NULL) {
1777 			nmp = mp_chain;
1778 			mp_chain = mp_chain->b_next;
1779 			freemsg(nmp);
1780 		}
1781 	}
1782 	/*
1783 	 * We're done here, so just free the original message and return the
1784 	 * new message chain, that could be NULL if failed, back to the caller.
1785 	 */
1786 	freemsg(mp);
1787 
1788 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1789 	    "<== nxge_do_softlso:mp_chain $%p", mp_chain));
1790 	return (mp_chain);
1791 }
1792 
1793 /*
1794  * Will be called before NIC driver do further operation on the message.
1795  * The input message may include LSO information, if so, go to softlso logic
1796  * to eliminate the oversized LSO packet for the incapable underlying h/w.
1797  * The return could be the same non-LSO message or a message chain for LSO case.
1798  *
1799  * The driver needs to call this function per packet and process the whole chain
1800  * if applied.
1801  */
1802 static mblk_t *
1803 nxge_lso_eliminate(mblk_t *mp)
1804 {
1805 	uint32_t lsoflags;
1806 	uint32_t mss;
1807 
1808 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1809 	    "==>nxge_lso_eliminate:"));
1810 	nxge_lso_info_get(mp, &mss, &lsoflags);
1811 
1812 	if (lsoflags & HW_LSO) {
1813 		mblk_t *nmp;
1814 
1815 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1816 		    "==>nxge_lso_eliminate:"
1817 		    "HW_LSO:mss %d mp $%p",
1818 		    mss, mp));
1819 		if ((nmp = nxge_do_softlso(mp, mss)) != NULL) {
1820 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1821 			    "<== nxge_lso_eliminate: "
1822 			    "LSO: nmp not NULL nmp $%p mss %d mp $%p",
1823 			    nmp, mss, mp));
1824 			return (nmp);
1825 		} else {
1826 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1827 			    "<== nxge_lso_eliminate_ "
1828 			    "LSO: failed nmp NULL nmp $%p mss %d mp $%p",
1829 			    nmp, mss, mp));
1830 			return (NULL);
1831 		}
1832 	}
1833 
1834 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1835 	    "<== nxge_lso_eliminate"));
1836 	return (mp);
1837 }
1838 
1839 static uint32_t
1840 nxge_csgen(uint16_t *adr, int len)
1841 {
1842 	int		i, odd;
1843 	uint32_t	sum = 0;
1844 	uint32_t	c = 0;
1845 
1846 	odd = len % 2;
1847 	for (i = 0; i < (len / 2); i++) {
1848 		sum += (adr[i] & 0xffff);
1849 	}
1850 	if (odd) {
1851 		sum += adr[len / 2] & 0xff00;
1852 	}
1853 	while ((c = ((sum & 0xffff0000) >> 16)) != 0) {
1854 		sum &= 0xffff;
1855 		sum += c;
1856 	}
1857 	return (~sum & 0xffff);
1858 }
1859