xref: /titanic_41/usr/src/uts/common/io/nxge/nxge_send.c (revision 2b637af7da499c08e98bb81902f8992f624e8d86)
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 	    (((int)tx_ring_p->tx_ring_size - (int)tx_ring_p->descs_pending) <
399 	    (int)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 				(void) cas32((uint32_t *)&tx_ring_p->queueing,
426 				    0, 1);
427 				tdc_stats->tx_no_desc++;
428 
429 				if (isLDOMservice(nxgep)) {
430 					tx_ring_p->tx_ring_busy = B_FALSE;
431 					if (tx_ring_p->tx_ring_offline) {
432 						(void) atomic_swap_32(
433 						    &tx_ring_p->tx_ring_offline,
434 						    NXGE_TX_RING_OFFLINED);
435 					}
436 				}
437 
438 				MUTEX_EXIT(&tx_ring_p->lock);
439 				status = 1;
440 				goto nxge_start_fail1;
441 			}
442 		}
443 	}
444 
445 	nmp = mp;
446 	i = sop_index = tx_ring_p->wr_index;
447 	nmblks = 0;
448 	ngathers = 0;
449 	pkt_len = 0;
450 	pack_len = 0;
451 	clen = 0;
452 	last_bidx = -1;
453 	good_packet = B_TRUE;
454 
455 	desc_area = tx_ring_p->tdc_desc;
456 	npi_handle = desc_area.npi_handle;
457 	npi_desc_handle.regh = (nxge_os_acc_handle_t)
458 	    DMA_COMMON_ACC_HANDLE(desc_area);
459 	tx_desc_ring_vp = (p_tx_desc_t)DMA_COMMON_VPTR(desc_area);
460 	tx_desc_dma_handle = (nxge_os_dma_handle_t)
461 	    DMA_COMMON_HANDLE(desc_area);
462 	tx_msg_ring = tx_ring_p->tx_msg_ring;
463 
464 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: wr_index %d i %d",
465 	    sop_index, i));
466 
467 #ifdef	NXGE_DEBUG
468 	msgsize = msgdsize(nmp);
469 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
470 	    "==> nxge_start(1): wr_index %d i %d msgdsize %d",
471 	    sop_index, i, msgsize));
472 #endif
473 	/*
474 	 * The first 16 bytes of the premapped buffer are reserved
475 	 * for header. No padding will be used.
476 	 */
477 	pkt_len = pack_len = boff = TX_PKT_HEADER_SIZE;
478 	if (nxge_tx_use_bcopy && (nxgep->niu_type != N2_NIU)) {
479 		bcopy_thresh = (nxge_bcopy_thresh - TX_PKT_HEADER_SIZE);
480 	} else {
481 		bcopy_thresh = (TX_BCOPY_SIZE - TX_PKT_HEADER_SIZE);
482 	}
483 	while (nmp) {
484 		good_packet = B_TRUE;
485 		b_rptr = nmp->b_rptr;
486 		len = MBLKL(nmp);
487 		if (len <= 0) {
488 			nmp = nmp->b_cont;
489 			continue;
490 		}
491 		nmblks++;
492 
493 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(1): nmblks %d "
494 		    "len %d pkt_len %d pack_len %d",
495 		    nmblks, len, pkt_len, pack_len));
496 		/*
497 		 * Hardware limits the transfer length to 4K for NIU and
498 		 * 4076 (TX_MAX_TRANSFER_LENGTH) for Neptune. But we just
499 		 * use TX_MAX_TRANSFER_LENGTH as the limit for both.
500 		 * If len is longer than the limit, then we break nmp into
501 		 * two chunks: Make the first chunk equal to the limit and
502 		 * the second chunk for the remaining data. If the second
503 		 * chunk is still larger than the limit, then it will be
504 		 * broken into two in the next pass.
505 		 */
506 		if (len > TX_MAX_TRANSFER_LENGTH - TX_PKT_HEADER_SIZE) {
507 			if ((t_mp = dupb(nmp)) != NULL) {
508 				nmp->b_wptr = nmp->b_rptr +
509 				    (TX_MAX_TRANSFER_LENGTH
510 				    - TX_PKT_HEADER_SIZE);
511 				t_mp->b_rptr = nmp->b_wptr;
512 				t_mp->b_cont = nmp->b_cont;
513 				nmp->b_cont = t_mp;
514 				len = MBLKL(nmp);
515 			} else {
516 				if (is_lso) {
517 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
518 					    "LSO packet: dupb failed: "
519 					    "channel %d",
520 					    tx_ring_p->tdc));
521 					mp = nmp;
522 					goto nxge_start_fail_lso;
523 				} else {
524 					good_packet = B_FALSE;
525 					goto nxge_start_fail2;
526 				}
527 			}
528 		}
529 		tx_desc.value = 0;
530 		tx_desc_p = &tx_desc_ring_vp[i];
531 #ifdef	NXGE_DEBUG
532 		tx_desc_pp = &tx_desc_ring_pp[i];
533 #endif
534 		tx_msg_p = &tx_msg_ring[i];
535 #if defined(__i386)
536 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
537 #else
538 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
539 #endif
540 		if (!header_set &&
541 		    ((!nxge_tx_use_bcopy && (len > TX_BCOPY_SIZE)) ||
542 		    (len >= bcopy_thresh))) {
543 			header_set = B_TRUE;
544 			bcopy_thresh += TX_PKT_HEADER_SIZE;
545 			boff = 0;
546 			pack_len = 0;
547 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
548 			hdrp = (p_tx_pkt_header_t)kaddr;
549 			clen = pkt_len;
550 			dma_handle = tx_msg_p->buf_dma_handle;
551 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
552 			(void) ddi_dma_sync(dma_handle,
553 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
554 			    DDI_DMA_SYNC_FORDEV);
555 
556 			tx_msg_p->flags.dma_type = USE_BCOPY;
557 			goto nxge_start_control_header_only;
558 		}
559 
560 		pkt_len += len;
561 		pack_len += len;
562 
563 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(3): "
564 		    "desc entry %d "
565 		    "DESC IOADDR $%p "
566 		    "desc_vp $%p tx_desc_p $%p "
567 		    "desc_pp $%p tx_desc_pp $%p "
568 		    "len %d pkt_len %d pack_len %d",
569 		    i,
570 		    DMA_COMMON_IOADDR(desc_area),
571 		    tx_desc_ring_vp, tx_desc_p,
572 		    tx_desc_ring_pp, tx_desc_pp,
573 		    len, pkt_len, pack_len));
574 
575 		if (len < bcopy_thresh) {
576 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(4): "
577 			    "USE BCOPY: "));
578 			if (nxge_tx_tiny_pack) {
579 				uint32_t blst =
580 				    TXDMA_DESC_NEXT_INDEX(i, -1,
581 				    tx_ring_p->tx_wrap_mask);
582 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
583 				    "==> nxge_start(5): pack"));
584 				if ((pack_len <= bcopy_thresh) &&
585 				    (last_bidx == blst)) {
586 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
587 					    "==> nxge_start: pack(6) "
588 					    "(pkt_len %d pack_len %d)",
589 					    pkt_len, pack_len));
590 					i = blst;
591 					tx_desc_p = &tx_desc_ring_vp[i];
592 #ifdef	NXGE_DEBUG
593 					tx_desc_pp = &tx_desc_ring_pp[i];
594 #endif
595 					tx_msg_p = &tx_msg_ring[i];
596 					boff = pack_len - len;
597 					ngathers--;
598 				} else if (pack_len > bcopy_thresh &&
599 				    header_set) {
600 					pack_len = len;
601 					boff = 0;
602 					bcopy_thresh = nxge_bcopy_thresh;
603 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
604 					    "==> nxge_start(7): > max NEW "
605 					    "bcopy thresh %d "
606 					    "pkt_len %d pack_len %d(next)",
607 					    bcopy_thresh,
608 					    pkt_len, pack_len));
609 				}
610 				last_bidx = i;
611 			}
612 			kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
613 			if ((boff == TX_PKT_HEADER_SIZE) && (nmblks == 1)) {
614 				hdrp = (p_tx_pkt_header_t)kaddr;
615 				header_set = B_TRUE;
616 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
617 				    "==> nxge_start(7_x2): "
618 				    "pkt_len %d pack_len %d (new hdrp $%p)",
619 				    pkt_len, pack_len, hdrp));
620 			}
621 			tx_msg_p->flags.dma_type = USE_BCOPY;
622 			kaddr += boff;
623 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(8): "
624 			    "USE BCOPY: before bcopy "
625 			    "DESC IOADDR $%p entry %d "
626 			    "bcopy packets %d "
627 			    "bcopy kaddr $%p "
628 			    "bcopy ioaddr (SAD) $%p "
629 			    "bcopy clen %d "
630 			    "bcopy boff %d",
631 			    DMA_COMMON_IOADDR(desc_area), i,
632 			    tdc_stats->tx_hdr_pkts,
633 			    kaddr,
634 			    dma_ioaddr,
635 			    clen,
636 			    boff));
637 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
638 			    "1USE BCOPY: "));
639 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
640 			    "2USE BCOPY: "));
641 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: "
642 			    "last USE BCOPY: copy from b_rptr $%p "
643 			    "to KADDR $%p (len %d offset %d",
644 			    b_rptr, kaddr, len, boff));
645 
646 			bcopy(b_rptr, kaddr, len);
647 
648 #ifdef	NXGE_DEBUG
649 			dump_len = (len > 128) ? 128: len;
650 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
651 			    "==> nxge_start: dump packets "
652 			    "(After BCOPY len %d)"
653 			    "(b_rptr $%p): %s", len, nmp->b_rptr,
654 			    nxge_dump_packet((char *)nmp->b_rptr,
655 			    dump_len)));
656 #endif
657 
658 			dma_handle = tx_msg_p->buf_dma_handle;
659 			dma_ioaddr = DMA_COMMON_IOADDR(tx_msg_p->buf_dma);
660 			(void) ddi_dma_sync(dma_handle,
661 			    i * nxge_bcopy_thresh, nxge_bcopy_thresh,
662 			    DDI_DMA_SYNC_FORDEV);
663 			clen = len + boff;
664 			tdc_stats->tx_hdr_pkts++;
665 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(9): "
666 			    "USE BCOPY: "
667 			    "DESC IOADDR $%p entry %d "
668 			    "bcopy packets %d "
669 			    "bcopy kaddr $%p "
670 			    "bcopy ioaddr (SAD) $%p "
671 			    "bcopy clen %d "
672 			    "bcopy boff %d",
673 			    DMA_COMMON_IOADDR(desc_area),
674 			    i,
675 			    tdc_stats->tx_hdr_pkts,
676 			    kaddr,
677 			    dma_ioaddr,
678 			    clen,
679 			    boff));
680 		} else {
681 			NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(12): "
682 			    "USE DVMA: len %d", len));
683 			tx_msg_p->flags.dma_type = USE_DMA;
684 			dma_flags = DDI_DMA_WRITE;
685 			if (len < nxge_dma_stream_thresh) {
686 				dma_flags |= DDI_DMA_CONSISTENT;
687 			} else {
688 				dma_flags |= DDI_DMA_STREAMING;
689 			}
690 
691 			dma_handle = tx_msg_p->dma_handle;
692 			status = ddi_dma_addr_bind_handle(dma_handle, NULL,
693 			    (caddr_t)b_rptr, len, dma_flags,
694 			    DDI_DMA_DONTWAIT, NULL,
695 			    &dma_cookie, &ncookies);
696 			if (status == DDI_DMA_MAPPED) {
697 				dma_ioaddr = dma_cookie.dmac_laddress;
698 				len = (int)dma_cookie.dmac_size;
699 				clen = (uint32_t)dma_cookie.dmac_size;
700 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
701 				    "==> nxge_start(12_1): "
702 				    "USE DVMA: len %d clen %d "
703 				    "ngathers %d",
704 				    len, clen,
705 				    ngathers));
706 #if defined(__i386)
707 				npi_desc_handle.regp = (uint32_t)tx_desc_p;
708 #else
709 				npi_desc_handle.regp = (uint64_t)tx_desc_p;
710 #endif
711 				while (ncookies > 1) {
712 					ngathers++;
713 					/*
714 					 * this is the fix for multiple
715 					 * cookies, which are basically
716 					 * a descriptor entry, we don't set
717 					 * SOP bit as well as related fields
718 					 */
719 
720 					(void) npi_txdma_desc_gather_set(
721 					    npi_desc_handle,
722 					    &tx_desc,
723 					    (ngathers -1),
724 					    mark_mode,
725 					    ngathers,
726 					    dma_ioaddr,
727 					    clen);
728 
729 					tx_msg_p->tx_msg_size = clen;
730 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
731 					    "==> nxge_start:  DMA "
732 					    "ncookie %d "
733 					    "ngathers %d "
734 					    "dma_ioaddr $%p len %d"
735 					    "desc $%p descp $%p (%d)",
736 					    ncookies,
737 					    ngathers,
738 					    dma_ioaddr, clen,
739 					    *tx_desc_p, tx_desc_p, i));
740 
741 					ddi_dma_nextcookie(dma_handle,
742 					    &dma_cookie);
743 					dma_ioaddr =
744 					    dma_cookie.dmac_laddress;
745 
746 					len = (int)dma_cookie.dmac_size;
747 					clen = (uint32_t)dma_cookie.dmac_size;
748 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
749 					    "==> nxge_start(12_2): "
750 					    "USE DVMA: len %d clen %d ",
751 					    len, clen));
752 
753 					i = TXDMA_DESC_NEXT_INDEX(i, 1,
754 					    tx_ring_p->tx_wrap_mask);
755 					tx_desc_p = &tx_desc_ring_vp[i];
756 
757 #if defined(__i386)
758 					npi_desc_handle.regp =
759 					    (uint32_t)tx_desc_p;
760 #else
761 					npi_desc_handle.regp =
762 					    (uint64_t)tx_desc_p;
763 #endif
764 					tx_msg_p = &tx_msg_ring[i];
765 					tx_msg_p->flags.dma_type = USE_NONE;
766 					tx_desc.value = 0;
767 
768 					ncookies--;
769 				}
770 				tdc_stats->tx_ddi_pkts++;
771 				NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start:"
772 				    "DMA: ddi packets %d",
773 				    tdc_stats->tx_ddi_pkts));
774 			} else {
775 				NXGE_ERROR_MSG((nxgep, NXGE_ERR_CTL,
776 				    "dma mapping failed for %d "
777 				    "bytes addr $%p flags %x (%d)",
778 				    len, b_rptr, status, status));
779 				good_packet = B_FALSE;
780 				tdc_stats->tx_dma_bind_fail++;
781 				tx_msg_p->flags.dma_type = USE_NONE;
782 				if (is_lso) {
783 					mp = nmp;
784 					goto nxge_start_fail_lso;
785 				} else {
786 					goto nxge_start_fail2;
787 				}
788 			}
789 		} /* ddi dvma */
790 
791 		if (is_lso) {
792 			nmp_lso_save = nmp;
793 		}
794 		nmp = nmp->b_cont;
795 nxge_start_control_header_only:
796 #if defined(__i386)
797 		npi_desc_handle.regp = (uint32_t)tx_desc_p;
798 #else
799 		npi_desc_handle.regp = (uint64_t)tx_desc_p;
800 #endif
801 		ngathers++;
802 
803 		if (ngathers == 1) {
804 #ifdef	NXGE_DEBUG
805 			save_desc_p = &sop_tx_desc;
806 #endif
807 			sop_tx_desc_p = &sop_tx_desc;
808 			sop_tx_desc_p->value = 0;
809 			sop_tx_desc_p->bits.hdw.tr_len = clen;
810 			sop_tx_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
811 			sop_tx_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
812 		} else {
813 #ifdef	NXGE_DEBUG
814 			save_desc_p = &tx_desc;
815 #endif
816 			tmp_desc_p = &tx_desc;
817 			tmp_desc_p->value = 0;
818 			tmp_desc_p->bits.hdw.tr_len = clen;
819 			tmp_desc_p->bits.hdw.sad = dma_ioaddr >> 32;
820 			tmp_desc_p->bits.ldw.sad = dma_ioaddr & 0xffffffff;
821 
822 			tx_desc_p->value = tmp_desc_p->value;
823 		}
824 
825 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(13): "
826 		    "Desc_entry %d ngathers %d "
827 		    "desc_vp $%p tx_desc_p $%p "
828 		    "len %d clen %d pkt_len %d pack_len %d nmblks %d "
829 		    "dma_ioaddr (SAD) $%p mark %d",
830 		    i, ngathers,
831 		    tx_desc_ring_vp, tx_desc_p,
832 		    len, clen, pkt_len, pack_len, nmblks,
833 		    dma_ioaddr, mark_mode));
834 
835 #ifdef NXGE_DEBUG
836 		npi_desc_handle.nxgep = nxgep;
837 		npi_desc_handle.function.function = nxgep->function_num;
838 		npi_desc_handle.function.instance = nxgep->instance;
839 		sad = (save_desc_p->value & TX_PKT_DESC_SAD_MASK);
840 		xfer_len = ((save_desc_p->value & TX_PKT_DESC_TR_LEN_MASK) >>
841 		    TX_PKT_DESC_TR_LEN_SHIFT);
842 
843 
844 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
845 		    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\t"
846 		    "mark %d sop %d\n",
847 		    save_desc_p->value,
848 		    sad,
849 		    save_desc_p->bits.hdw.tr_len,
850 		    xfer_len,
851 		    save_desc_p->bits.hdw.num_ptr,
852 		    save_desc_p->bits.hdw.mark,
853 		    save_desc_p->bits.hdw.sop));
854 
855 		npi_txdma_dump_desc_one(npi_desc_handle, NULL, i);
856 #endif
857 
858 		tx_msg_p->tx_msg_size = clen;
859 		i = TXDMA_DESC_NEXT_INDEX(i, 1, tx_ring_p->tx_wrap_mask);
860 		if (ngathers > nxge_tx_max_gathers) {
861 			good_packet = B_FALSE;
862 			hcksum_retrieve(mp, NULL, NULL, &start_offset,
863 			    &stuff_offset, &end_offset, &value,
864 			    &cksum_flags);
865 
866 			NXGE_DEBUG_MSG((NULL, TX_CTL,
867 			    "==> nxge_start(14): pull msg - "
868 			    "len %d pkt_len %d ngathers %d",
869 			    len, pkt_len, ngathers));
870 			/* Pull all message blocks from b_cont */
871 			if (is_lso) {
872 				mp = nmp_lso_save;
873 				goto nxge_start_fail_lso;
874 			}
875 			if ((msgpullup(mp, -1)) == NULL) {
876 				goto nxge_start_fail2;
877 			}
878 			goto nxge_start_fail2;
879 		}
880 	} /* while (nmp) */
881 
882 	tx_msg_p->tx_message = mp;
883 	tx_desc_p = &tx_desc_ring_vp[sop_index];
884 #if defined(__i386)
885 	npi_desc_handle.regp = (uint32_t)tx_desc_p;
886 #else
887 	npi_desc_handle.regp = (uint64_t)tx_desc_p;
888 #endif
889 
890 	pkthdrp = (p_tx_pkt_hdr_all_t)hdrp;
891 	pkthdrp->reserved = 0;
892 	hdrp->value = 0;
893 	bcopy(&tmp_hdrp, hdrp, sizeof (tx_pkt_header_t));
894 
895 	if (pkt_len > NXGE_MTU_DEFAULT_MAX) {
896 		tdc_stats->tx_jumbo_pkts++;
897 	}
898 
899 	min_len = (ETHERMIN + TX_PKT_HEADER_SIZE + (npads * 2));
900 	if (pkt_len < min_len) {
901 		/* Assume we use bcopy to premapped buffers */
902 		kaddr = (caddr_t)DMA_COMMON_VPTR(tx_msg_p->buf_dma);
903 		NXGE_DEBUG_MSG((NULL, TX_CTL,
904 		    "==> nxge_start(14-1): < (msg_min + 16)"
905 		    "len %d pkt_len %d min_len %d bzero %d ngathers %d",
906 		    len, pkt_len, min_len, (min_len - pkt_len), ngathers));
907 		bzero((kaddr + pkt_len), (min_len - pkt_len));
908 		pkt_len = tx_msg_p->tx_msg_size = min_len;
909 
910 		sop_tx_desc_p->bits.hdw.tr_len = min_len;
911 
912 		NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
913 		tx_desc_p->value = sop_tx_desc_p->value;
914 
915 		NXGE_DEBUG_MSG((NULL, TX_CTL,
916 		    "==> nxge_start(14-2): < msg_min - "
917 		    "len %d pkt_len %d min_len %d ngathers %d",
918 		    len, pkt_len, min_len, ngathers));
919 	}
920 
921 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: cksum_flags 0x%x ",
922 	    cksum_flags));
923 	{
924 		uint64_t	tmp_len;
925 
926 		/* pkt_len already includes 16 + paddings!! */
927 		/* Update the control header length */
928 		tot_xfer_len = (pkt_len - TX_PKT_HEADER_SIZE);
929 		tmp_len = hdrp->value |
930 		    (tot_xfer_len << TX_PKT_HEADER_TOT_XFER_LEN_SHIFT);
931 
932 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
933 		    "==> nxge_start(15_x1): setting SOP "
934 		    "tot_xfer_len 0x%llx (%d) pkt_len %d tmp_len "
935 		    "0x%llx hdrp->value 0x%llx",
936 		    tot_xfer_len, tot_xfer_len, pkt_len,
937 		    tmp_len, hdrp->value));
938 #if defined(_BIG_ENDIAN)
939 		hdrp->value = ddi_swap64(tmp_len);
940 #else
941 		hdrp->value = tmp_len;
942 #endif
943 		NXGE_DEBUG_MSG((nxgep,
944 		    TX_CTL, "==> nxge_start(15_x2): setting SOP "
945 		    "after SWAP: tot_xfer_len 0x%llx pkt_len %d "
946 		    "tmp_len 0x%llx hdrp->value 0x%llx",
947 		    tot_xfer_len, pkt_len,
948 		    tmp_len, hdrp->value));
949 	}
950 
951 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(15): setting SOP "
952 	    "wr_index %d "
953 	    "tot_xfer_len (%d) pkt_len %d npads %d",
954 	    sop_index,
955 	    tot_xfer_len, pkt_len,
956 	    npads));
957 
958 	sop_tx_desc_p->bits.hdw.sop = 1;
959 	sop_tx_desc_p->bits.hdw.mark = mark_mode;
960 	sop_tx_desc_p->bits.hdw.num_ptr = ngathers;
961 
962 	NXGE_MEM_PIO_WRITE64(npi_desc_handle, sop_tx_desc_p->value);
963 
964 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start(16): set SOP done"));
965 
966 #ifdef NXGE_DEBUG
967 	npi_desc_handle.nxgep = nxgep;
968 	npi_desc_handle.function.function = nxgep->function_num;
969 	npi_desc_handle.function.instance = nxgep->instance;
970 
971 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "\n\t: value 0x%llx\n"
972 	    "\t\tsad $%p\ttr_len %d len %d\tnptrs %d\tmark %d sop %d\n",
973 	    save_desc_p->value,
974 	    sad,
975 	    save_desc_p->bits.hdw.tr_len,
976 	    xfer_len,
977 	    save_desc_p->bits.hdw.num_ptr,
978 	    save_desc_p->bits.hdw.mark,
979 	    save_desc_p->bits.hdw.sop));
980 	(void) npi_txdma_dump_desc_one(npi_desc_handle, NULL, sop_index);
981 
982 	dump_len = (pkt_len > 128) ? 128: pkt_len;
983 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
984 	    "==> nxge_start: dump packets(17) (after sop set, len "
985 	    " (len/dump_len/pkt_len/tot_xfer_len) %d/%d/%d/%d):\n"
986 	    "ptr $%p: %s", len, dump_len, pkt_len, tot_xfer_len,
987 	    (char *)hdrp,
988 	    nxge_dump_packet((char *)hdrp, dump_len)));
989 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
990 	    "==> nxge_start(18): TX desc sync: sop_index %d",
991 	    sop_index));
992 #endif
993 
994 	if ((ngathers == 1) || tx_ring_p->wr_index < i) {
995 		(void) ddi_dma_sync(tx_desc_dma_handle,
996 		    sop_index * sizeof (tx_desc_t),
997 		    ngathers * sizeof (tx_desc_t),
998 		    DDI_DMA_SYNC_FORDEV);
999 
1000 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(19): sync 1 "
1001 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1002 		    "pkt_len %d ngathers %d sop_index %d\n",
1003 		    stuff_offset, start_offset,
1004 		    pkt_len, ngathers, sop_index));
1005 	} else { /* more than one descriptor and wrap around */
1006 		uint32_t nsdescs = tx_ring_p->tx_ring_size - sop_index;
1007 		(void) ddi_dma_sync(tx_desc_dma_handle,
1008 		    sop_index * sizeof (tx_desc_t),
1009 		    nsdescs * sizeof (tx_desc_t),
1010 		    DDI_DMA_SYNC_FORDEV);
1011 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(20): sync 1 "
1012 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1013 		    "pkt_len %d ngathers %d sop_index %d\n",
1014 		    stuff_offset, start_offset,
1015 		    pkt_len, ngathers, sop_index));
1016 
1017 		(void) ddi_dma_sync(tx_desc_dma_handle,
1018 		    0,
1019 		    (ngathers - nsdescs) * sizeof (tx_desc_t),
1020 		    DDI_DMA_SYNC_FORDEV);
1021 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "nxge_start(21): sync 2 "
1022 		    "cs_off = 0x%02X cs_s_off = 0x%02X "
1023 		    "pkt_len %d ngathers %d sop_index %d\n",
1024 		    stuff_offset, start_offset,
1025 		    pkt_len, ngathers, sop_index));
1026 	}
1027 
1028 	tail_index = tx_ring_p->wr_index;
1029 	tail_wrap = tx_ring_p->wr_index_wrap;
1030 
1031 	tx_ring_p->wr_index = i;
1032 	if (tx_ring_p->wr_index <= tail_index) {
1033 		tx_ring_p->wr_index_wrap = ((tail_wrap == B_TRUE) ?
1034 		    B_FALSE : B_TRUE);
1035 	}
1036 
1037 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX kick: "
1038 	    "channel %d wr_index %d wrap %d ngathers %d desc_pend %d",
1039 	    tx_ring_p->tdc,
1040 	    tx_ring_p->wr_index,
1041 	    tx_ring_p->wr_index_wrap,
1042 	    ngathers,
1043 	    tx_ring_p->descs_pending));
1044 
1045 	if (is_lso) {
1046 		lso_ngathers += ngathers;
1047 		if (mp_chain != NULL) {
1048 			mp = mp_chain;
1049 			mp_chain = mp_chain->b_next;
1050 			mp->b_next = NULL;
1051 			if (nxge_lso_kick_cnt == lso_ngathers) {
1052 				tx_ring_p->descs_pending += lso_ngathers;
1053 				{
1054 					tx_ring_kick_t		kick;
1055 
1056 					kick.value = 0;
1057 					kick.bits.ldw.wrap =
1058 					    tx_ring_p->wr_index_wrap;
1059 					kick.bits.ldw.tail =
1060 					    (uint16_t)tx_ring_p->wr_index;
1061 
1062 					/* Kick the Transmit kick register */
1063 					TXDMA_REG_WRITE64(
1064 					    NXGE_DEV_NPI_HANDLE(nxgep),
1065 					    TX_RING_KICK_REG,
1066 					    (uint8_t)tx_ring_p->tdc,
1067 					    kick.value);
1068 					tdc_stats->tx_starts++;
1069 
1070 					NXGE_DEBUG_MSG((nxgep, TX_CTL,
1071 					    "==> nxge_start: more LSO: "
1072 					    "LSO_CNT %d",
1073 					    lso_ngathers));
1074 				}
1075 				lso_ngathers = 0;
1076 				ngathers = 0;
1077 				cur_index_lso = sop_index = tx_ring_p->wr_index;
1078 				lso_tail_wrap = tx_ring_p->wr_index_wrap;
1079 			}
1080 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1081 			    "==> nxge_start: lso again: "
1082 			    "lso_gathers %d ngathers %d cur_index_lso %d "
1083 			    "wr_index %d sop_index %d",
1084 			    lso_ngathers, ngathers, cur_index_lso,
1085 			    tx_ring_p->wr_index, sop_index));
1086 
1087 			NXGE_DEBUG_MSG((nxgep, TX_CTL,
1088 			    "==> nxge_start: next : count %d",
1089 			    lso_ngathers));
1090 			lso_again = B_TRUE;
1091 			goto start_again;
1092 		}
1093 		ngathers = lso_ngathers;
1094 	}
1095 
1096 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: TX KICKING: "));
1097 
1098 	{
1099 		tx_ring_kick_t		kick;
1100 
1101 		kick.value = 0;
1102 		kick.bits.ldw.wrap = tx_ring_p->wr_index_wrap;
1103 		kick.bits.ldw.tail = (uint16_t)tx_ring_p->wr_index;
1104 
1105 		/* Kick start the Transmit kick register */
1106 		TXDMA_REG_WRITE64(NXGE_DEV_NPI_HANDLE(nxgep),
1107 		    TX_RING_KICK_REG,
1108 		    (uint8_t)tx_ring_p->tdc,
1109 		    kick.value);
1110 	}
1111 
1112 	tx_ring_p->descs_pending += ngathers;
1113 	tdc_stats->tx_starts++;
1114 
1115 	if (isLDOMservice(nxgep)) {
1116 		tx_ring_p->tx_ring_busy = B_FALSE;
1117 		if (tx_ring_p->tx_ring_offline) {
1118 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1119 			    NXGE_TX_RING_OFFLINED);
1120 		}
1121 	}
1122 
1123 	MUTEX_EXIT(&tx_ring_p->lock);
1124 
1125 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1126 	return (status);
1127 
1128 nxge_start_fail_lso:
1129 	status = 0;
1130 	good_packet = B_FALSE;
1131 	if (mp != NULL) {
1132 		freemsg(mp);
1133 	}
1134 	if (mp_chain != NULL) {
1135 		freemsg(mp_chain);
1136 	}
1137 	if (!lso_again && !ngathers) {
1138 		if (isLDOMservice(nxgep)) {
1139 			tx_ring_p->tx_ring_busy = B_FALSE;
1140 			if (tx_ring_p->tx_ring_offline) {
1141 				(void) atomic_swap_32(
1142 				    &tx_ring_p->tx_ring_offline,
1143 				    NXGE_TX_RING_OFFLINED);
1144 			}
1145 		}
1146 
1147 		MUTEX_EXIT(&tx_ring_p->lock);
1148 		NXGE_DEBUG_MSG((nxgep, TX_CTL,
1149 		    "==> nxge_start: lso exit (nothing changed)"));
1150 		goto nxge_start_fail1;
1151 	}
1152 
1153 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1154 	    "==> nxge_start (channel %d): before lso "
1155 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1156 	    "wr_index %d sop_index %d lso_again %d",
1157 	    tx_ring_p->tdc,
1158 	    lso_ngathers, ngathers, cur_index_lso,
1159 	    tx_ring_p->wr_index, sop_index, lso_again));
1160 
1161 	if (lso_again) {
1162 		lso_ngathers += ngathers;
1163 		ngathers = lso_ngathers;
1164 		sop_index = cur_index_lso;
1165 		tx_ring_p->wr_index = sop_index;
1166 		tx_ring_p->wr_index_wrap = lso_tail_wrap;
1167 	}
1168 
1169 	NXGE_DEBUG_MSG((nxgep, TX_CTL,
1170 	    "==> nxge_start (channel %d): after lso "
1171 	    "lso_gathers %d ngathers %d cur_index_lso %d "
1172 	    "wr_index %d sop_index %d lso_again %d",
1173 	    tx_ring_p->tdc,
1174 	    lso_ngathers, ngathers, cur_index_lso,
1175 	    tx_ring_p->wr_index, sop_index, lso_again));
1176 
1177 nxge_start_fail2:
1178 	if (good_packet == B_FALSE) {
1179 		cur_index = sop_index;
1180 		NXGE_DEBUG_MSG((nxgep, TX_CTL, "==> nxge_start: clean up"));
1181 		for (i = 0; i < ngathers; i++) {
1182 			tx_desc_p = &tx_desc_ring_vp[cur_index];
1183 #if defined(__i386)
1184 			npi_handle.regp = (uint32_t)tx_desc_p;
1185 #else
1186 			npi_handle.regp = (uint64_t)tx_desc_p;
1187 #endif
1188 			tx_msg_p = &tx_msg_ring[cur_index];
1189 			(void) npi_txdma_desc_set_zero(npi_handle, 1);
1190 			if (tx_msg_p->flags.dma_type == USE_DVMA) {
1191 				NXGE_DEBUG_MSG((nxgep, TX_CTL,
1192 				    "tx_desc_p = %X index = %d",
1193 				    tx_desc_p, tx_ring_p->rd_index));
1194 				(void) dvma_unload(tx_msg_p->dvma_handle,
1195 				    0, -1);
1196 				tx_msg_p->dvma_handle = NULL;
1197 				if (tx_ring_p->dvma_wr_index ==
1198 				    tx_ring_p->dvma_wrap_mask)
1199 					tx_ring_p->dvma_wr_index = 0;
1200 				else
1201 					tx_ring_p->dvma_wr_index++;
1202 				tx_ring_p->dvma_pending--;
1203 			} else if (tx_msg_p->flags.dma_type == USE_DMA) {
1204 				if (ddi_dma_unbind_handle(
1205 				    tx_msg_p->dma_handle)) {
1206 					cmn_err(CE_WARN, "!nxge_start: "
1207 					    "ddi_dma_unbind_handle failed");
1208 				}
1209 			}
1210 			tx_msg_p->flags.dma_type = USE_NONE;
1211 			cur_index = TXDMA_DESC_NEXT_INDEX(cur_index, 1,
1212 			    tx_ring_p->tx_wrap_mask);
1213 
1214 		}
1215 	}
1216 
1217 	if (isLDOMservice(nxgep)) {
1218 		tx_ring_p->tx_ring_busy = B_FALSE;
1219 		if (tx_ring_p->tx_ring_offline) {
1220 			(void) atomic_swap_32(&tx_ring_p->tx_ring_offline,
1221 			    NXGE_TX_RING_OFFLINED);
1222 		}
1223 	}
1224 
1225 	MUTEX_EXIT(&tx_ring_p->lock);
1226 
1227 nxge_start_fail1:
1228 	/* Add FMA to check the access handle nxge_hregh */
1229 
1230 	NXGE_DEBUG_MSG((nxgep, TX_CTL, "<== nxge_start"));
1231 	return (status);
1232 }
1233 
1234 /* Software LSO starts here */
1235 static void
1236 nxge_hcksum_retrieve(mblk_t *mp,
1237     uint32_t *start, uint32_t *stuff, uint32_t *end,
1238     uint32_t *value, uint32_t *flags)
1239 {
1240 	if (mp->b_datap->db_type == M_DATA) {
1241 		if (flags != NULL) {
1242 			*flags = DB_CKSUMFLAGS(mp) & (HCK_IPV4_HDRCKSUM |
1243 			    HCK_PARTIALCKSUM | HCK_FULLCKSUM |
1244 			    HCK_FULLCKSUM_OK);
1245 			if ((*flags & (HCK_PARTIALCKSUM |
1246 			    HCK_FULLCKSUM)) != 0) {
1247 				if (value != NULL)
1248 					*value = (uint32_t)DB_CKSUM16(mp);
1249 				if ((*flags & HCK_PARTIALCKSUM) != 0) {
1250 					if (start != NULL)
1251 						*start =
1252 						    (uint32_t)DB_CKSUMSTART(mp);
1253 					if (stuff != NULL)
1254 						*stuff =
1255 						    (uint32_t)DB_CKSUMSTUFF(mp);
1256 					if (end != NULL)
1257 						*end =
1258 						    (uint32_t)DB_CKSUMEND(mp);
1259 				}
1260 			}
1261 		}
1262 	}
1263 }
1264 
1265 static void
1266 nxge_lso_info_get(mblk_t *mp, uint32_t *mss, uint32_t *flags)
1267 {
1268 	ASSERT(DB_TYPE(mp) == M_DATA);
1269 
1270 	*mss = 0;
1271 	if (flags != NULL) {
1272 		*flags = DB_CKSUMFLAGS(mp) & HW_LSO;
1273 		if ((*flags != 0) && (mss != NULL)) {
1274 			*mss = (uint32_t)DB_LSOMSS(mp);
1275 		}
1276 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1277 		    "==> nxge_lso_info_get(flag !=NULL): mss %d *flags 0x%x",
1278 		    *mss, *flags));
1279 	}
1280 
1281 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1282 	    "<== nxge_lso_info_get: mss %d", *mss));
1283 }
1284 
1285 /*
1286  * Do Soft LSO on the oversized packet.
1287  *
1288  * 1. Create a chain of message for headers.
1289  * 2. Fill up header messages with proper information.
1290  * 3. Copy Eithernet, IP, and TCP headers from the original message to
1291  *    each new message with necessary adjustments.
1292  *    * Unchange the ethernet header for DIX frames. (by default)
1293  *    * IP Total Length field is updated to MSS or less(only for the last one).
1294  *    * IP Identification value is incremented by one for each packet.
1295  *    * TCP sequence Number is recalculated according to the payload length.
1296  *    * Set FIN and/or PSH flags for the *last* packet if applied.
1297  *    * TCP partial Checksum
1298  * 4. Update LSO information in the first message header.
1299  * 5. Release the original message header.
1300  */
1301 static mblk_t *
1302 nxge_do_softlso(mblk_t *mp, uint32_t mss)
1303 {
1304 	uint32_t	hckflags;
1305 	int		pktlen;
1306 	int		hdrlen;
1307 	int		segnum;
1308 	int		i;
1309 	struct ether_vlan_header *evh;
1310 	int		ehlen, iphlen, tcphlen;
1311 	struct ip	*oiph, *niph;
1312 	struct tcphdr *otcph, *ntcph;
1313 	int		available, len, left;
1314 	uint16_t	ip_id;
1315 	uint32_t	tcp_seq;
1316 #ifdef __sparc
1317 	uint32_t	tcp_seq_tmp;
1318 #endif
1319 	mblk_t		*datamp;
1320 	uchar_t		*rptr;
1321 	mblk_t		*nmp;
1322 	mblk_t		*cmp;
1323 	mblk_t		*mp_chain;
1324 	boolean_t do_cleanup = B_FALSE;
1325 	t_uscalar_t start_offset = 0;
1326 	t_uscalar_t stuff_offset = 0;
1327 	t_uscalar_t value = 0;
1328 	uint16_t	l4_len;
1329 	ipaddr_t	src, dst;
1330 	uint32_t	cksum, sum, l4cksum;
1331 
1332 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1333 	    "==> nxge_do_softlso"));
1334 	/*
1335 	 * check the length of LSO packet payload and calculate the number of
1336 	 * segments to be generated.
1337 	 */
1338 	pktlen = msgsize(mp);
1339 	evh = (struct ether_vlan_header *)mp->b_rptr;
1340 
1341 	/* VLAN? */
1342 	if (evh->ether_tpid == htons(ETHERTYPE_VLAN))
1343 		ehlen = sizeof (struct ether_vlan_header);
1344 	else
1345 		ehlen = sizeof (struct ether_header);
1346 	oiph = (struct ip *)(mp->b_rptr + ehlen);
1347 	iphlen = oiph->ip_hl * 4;
1348 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1349 	tcphlen = otcph->th_off * 4;
1350 
1351 	l4_len = pktlen - ehlen - iphlen;
1352 
1353 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1354 	    "==> nxge_do_softlso: mss %d oiph $%p "
1355 	    "original ip_sum oiph->ip_sum 0x%x "
1356 	    "original tcp_sum otcph->th_sum 0x%x "
1357 	    "oiph->ip_len %d pktlen %d ehlen %d "
1358 	    "l4_len %d (0x%x) ip_len - iphlen %d ",
1359 	    mss,
1360 	    oiph,
1361 	    oiph->ip_sum,
1362 	    otcph->th_sum,
1363 	    ntohs(oiph->ip_len), pktlen,
1364 	    ehlen,
1365 	    l4_len,
1366 	    l4_len,
1367 	    ntohs(oiph->ip_len) - iphlen));
1368 
1369 	/* IPv4 + TCP */
1370 	if (!(oiph->ip_v == IPV4_VERSION)) {
1371 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1372 		    "<== nxge_do_softlso: not IPV4 "
1373 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1374 		    ntohs(oiph->ip_len), pktlen, ehlen,
1375 		    tcphlen));
1376 		freemsg(mp);
1377 		return (NULL);
1378 	}
1379 
1380 	if (!(oiph->ip_p == IPPROTO_TCP)) {
1381 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1382 		    "<== nxge_do_softlso: not TCP "
1383 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1384 		    ntohs(oiph->ip_len), pktlen, ehlen,
1385 		    tcphlen));
1386 		freemsg(mp);
1387 		return (NULL);
1388 	}
1389 
1390 	if (!(ntohs(oiph->ip_len) == pktlen - ehlen)) {
1391 		NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1392 		    "<== nxge_do_softlso: len not matched  "
1393 		    "oiph->ip_len %d pktlen %d ehlen %d tcphlen %d",
1394 		    ntohs(oiph->ip_len), pktlen, ehlen,
1395 		    tcphlen));
1396 		freemsg(mp);
1397 		return (NULL);
1398 	}
1399 
1400 	otcph = (struct tcphdr *)(mp->b_rptr + ehlen + iphlen);
1401 	tcphlen = otcph->th_off * 4;
1402 
1403 	/* TCP flags can not include URG, RST, or SYN */
1404 	VERIFY((otcph->th_flags & (TH_SYN | TH_RST | TH_URG)) == 0);
1405 
1406 	hdrlen = ehlen + iphlen + tcphlen;
1407 
1408 	VERIFY(MBLKL(mp) >= hdrlen);
1409 
1410 	if (MBLKL(mp) > hdrlen) {
1411 		datamp = mp;
1412 		rptr = mp->b_rptr + hdrlen;
1413 	} else { /* = */
1414 		datamp = mp->b_cont;
1415 		rptr = datamp->b_rptr;
1416 	}
1417 
1418 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1419 	    "nxge_do_softlso: otcph $%p pktlen: %d, "
1420 	    "hdrlen %d ehlen %d iphlen %d tcphlen %d "
1421 	    "mblkl(mp): %d, mblkl(datamp): %d",
1422 	    otcph,
1423 	    pktlen, hdrlen, ehlen, iphlen, tcphlen,
1424 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1425 
1426 	hckflags = 0;
1427 	nxge_hcksum_retrieve(mp,
1428 	    &start_offset, &stuff_offset, &value, NULL, &hckflags);
1429 
1430 	dst = oiph->ip_dst.s_addr;
1431 	src = oiph->ip_src.s_addr;
1432 
1433 	cksum = (dst >> 16) + (dst & 0xFFFF) +
1434 	    (src >> 16) + (src & 0xFFFF);
1435 	l4cksum = cksum + IP_TCP_CSUM_COMP;
1436 
1437 	sum = l4_len + l4cksum;
1438 	sum = (sum & 0xFFFF) + (sum >> 16);
1439 
1440 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1441 	    "==> nxge_do_softlso: dst 0x%x src 0x%x sum 0x%x ~new 0x%x "
1442 	    "hckflags 0x%x start_offset %d stuff_offset %d "
1443 	    "value (original) 0x%x th_sum 0x%x "
1444 	    "pktlen %d l4_len %d (0x%x) "
1445 	    "MBLKL(mp): %d, MBLKL(datamp): %d dump header %s",
1446 	    dst, src,
1447 	    (sum & 0xffff), (~sum & 0xffff),
1448 	    hckflags, start_offset, stuff_offset,
1449 	    value, otcph->th_sum,
1450 	    pktlen,
1451 	    l4_len,
1452 	    l4_len,
1453 	    ntohs(oiph->ip_len) - (int)MBLKL(mp),
1454 	    (int)MBLKL(datamp),
1455 	    nxge_dump_packet((char *)evh, 12)));
1456 
1457 	/*
1458 	 * Start to process.
1459 	 */
1460 	available = pktlen - hdrlen;
1461 	segnum = (available - 1) / mss + 1;
1462 
1463 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1464 	    "==> nxge_do_softlso: pktlen %d "
1465 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1466 	    "available %d mss %d segnum %d",
1467 	    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp),
1468 	    available,
1469 	    mss,
1470 	    segnum));
1471 
1472 	VERIFY(segnum >= 2);
1473 
1474 	/*
1475 	 * Try to pre-allocate all header messages
1476 	 */
1477 	mp_chain = NULL;
1478 	for (i = 0; i < segnum; i++) {
1479 		if ((nmp = allocb(hdrlen, 0)) == NULL) {
1480 			/* Clean up the mp_chain */
1481 			while (mp_chain != NULL) {
1482 				nmp = mp_chain;
1483 				mp_chain = mp_chain->b_next;
1484 				freemsg(nmp);
1485 			}
1486 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1487 			    "<== nxge_do_softlso: "
1488 			    "Could not allocate enough messages for headers!"));
1489 			freemsg(mp);
1490 			return (NULL);
1491 		}
1492 		nmp->b_next = mp_chain;
1493 		mp_chain = nmp;
1494 
1495 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1496 		    "==> nxge_do_softlso: "
1497 		    "mp $%p nmp $%p mp_chain $%p mp_chain->b_next $%p",
1498 		    mp, nmp, mp_chain, mp_chain->b_next));
1499 	}
1500 
1501 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1502 	    "==> nxge_do_softlso: mp $%p nmp $%p mp_chain $%p",
1503 	    mp, nmp, mp_chain));
1504 
1505 	/*
1506 	 * Associate payload with new packets
1507 	 */
1508 	cmp = mp_chain;
1509 	left = available;
1510 	while (cmp != NULL) {
1511 		nmp = dupb(datamp);
1512 		if (nmp == NULL) {
1513 			do_cleanup = B_TRUE;
1514 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1515 			    "==>nxge_do_softlso: "
1516 			    "Can not dupb(datamp), have to do clean up"));
1517 			goto cleanup_allocated_msgs;
1518 		}
1519 
1520 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1521 		    "==> nxge_do_softlso: (loop) before mp $%p cmp $%p "
1522 		    "dupb nmp $%p len %d left %d msd %d ",
1523 		    mp, cmp, nmp, len, left, mss));
1524 
1525 		cmp->b_cont = nmp;
1526 		nmp->b_rptr = rptr;
1527 		len = (left < mss) ? left : mss;
1528 		left -= len;
1529 
1530 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1531 		    "==> nxge_do_softlso: (loop) after mp $%p cmp $%p "
1532 		    "dupb nmp $%p len %d left %d mss %d ",
1533 		    mp, cmp, nmp, len, left, mss));
1534 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1535 		    "nxge_do_softlso: before available: %d, "
1536 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1537 		    available, left, len, segnum, (int)MBLKL(nmp)));
1538 
1539 		len -= MBLKL(nmp);
1540 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1541 		    "nxge_do_softlso: after available: %d, "
1542 		    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1543 		    available, left, len, segnum, (int)MBLKL(nmp)));
1544 
1545 		while (len > 0) {
1546 			mblk_t *mmp = NULL;
1547 
1548 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1549 			    "nxge_do_softlso: (4) len > 0 available: %d, "
1550 			    "left: %d, len: %d, segnum: %d MBLK(nmp): %d",
1551 			    available, left, len, segnum, (int)MBLKL(nmp)));
1552 
1553 			if (datamp->b_cont != NULL) {
1554 				datamp = datamp->b_cont;
1555 				rptr = datamp->b_rptr;
1556 				mmp = dupb(datamp);
1557 				if (mmp == NULL) {
1558 					do_cleanup = B_TRUE;
1559 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1560 					    "==> nxge_do_softlso: "
1561 					    "Can not dupb(datamp) (1), :"
1562 					    "have to do clean up"));
1563 					NXGE_DEBUG_MSG((NULL, TX_CTL,
1564 					    "==> nxge_do_softlso: "
1565 					    "available: %d, left: %d, "
1566 					    "len: %d, MBLKL(nmp): %d",
1567 					    available, left, len,
1568 					    (int)MBLKL(nmp)));
1569 					goto cleanup_allocated_msgs;
1570 				}
1571 			} else {
1572 				NXGE_ERROR_MSG((NULL, NXGE_ERR_CTL,
1573 				    "==> nxge_do_softlso: "
1574 				    "(1)available: %d, left: %d, "
1575 				    "len: %d, MBLKL(nmp): %d",
1576 				    available, left, len,
1577 				    (int)MBLKL(nmp)));
1578 				cmn_err(CE_PANIC,
1579 				    "==> nxge_do_softlso: "
1580 				    "Pointers must have been corrupted!\n"
1581 				    "datamp: $%p, nmp: $%p, rptr: $%p",
1582 				    (void *)datamp,
1583 				    (void *)nmp,
1584 				    (void *)rptr);
1585 			}
1586 			nmp->b_cont = mmp;
1587 			nmp = mmp;
1588 			len -= MBLKL(nmp);
1589 		}
1590 		if (len < 0) {
1591 			nmp->b_wptr += len;
1592 			rptr = nmp->b_wptr;
1593 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1594 			    "(5) len < 0 (less than 0)"
1595 			    "available: %d, left: %d, len: %d, MBLKL(nmp): %d",
1596 			    available, left, len, (int)MBLKL(nmp)));
1597 
1598 		} else if (len == 0) {
1599 			if (datamp->b_cont != NULL) {
1600 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1601 				    "(5) len == 0"
1602 				    "available: %d, left: %d, len: %d, "
1603 				    "MBLKL(nmp): %d",
1604 				    available, left, len, (int)MBLKL(nmp)));
1605 				datamp = datamp->b_cont;
1606 				rptr = datamp->b_rptr;
1607 			} else {
1608 				NXGE_DEBUG_MSG((NULL, TX_CTL,
1609 				    "(6)available b_cont == NULL : %d, "
1610 				    "left: %d, len: %d, MBLKL(nmp): %d",
1611 				    available, left, len, (int)MBLKL(nmp)));
1612 
1613 				VERIFY(cmp->b_next == NULL);
1614 				VERIFY(left == 0);
1615 				break; /* Done! */
1616 			}
1617 		}
1618 		cmp = cmp->b_next;
1619 
1620 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1621 		    "(7) do_softlso: "
1622 		    "next mp in mp_chain available len != 0 : %d, "
1623 		    "left: %d, len: %d, MBLKL(nmp): %d",
1624 		    available, left, len, (int)MBLKL(nmp)));
1625 	}
1626 
1627 	/*
1628 	 * From now, start to fill up all headers for the first message
1629 	 * Hardware checksum flags need to be updated separately for FULLCKSUM
1630 	 * and PARTIALCKSUM cases. For full checksum, copy the original flags
1631 	 * into every new packet is enough. But for HCK_PARTIALCKSUM, all
1632 	 * required fields need to be updated properly.
1633 	 */
1634 	nmp = mp_chain;
1635 	bcopy(mp->b_rptr, nmp->b_rptr, hdrlen);
1636 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1637 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1638 	niph->ip_len = htons(mss + iphlen + tcphlen);
1639 	ip_id = ntohs(niph->ip_id);
1640 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1641 #ifdef __sparc
1642 	bcopy((char *)&ntcph->th_seq, &tcp_seq_tmp, 4);
1643 	tcp_seq = ntohl(tcp_seq_tmp);
1644 #else
1645 	tcp_seq = ntohl(ntcph->th_seq);
1646 #endif
1647 
1648 	ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST);
1649 
1650 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1651 	DB_CKSUMSTART(nmp) = start_offset;
1652 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1653 
1654 	/* calculate IP checksum and TCP pseudo header checksum */
1655 	niph->ip_sum = 0;
1656 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1657 
1658 	l4_len = mss + tcphlen;
1659 	sum = htons(l4_len) + l4cksum;
1660 	sum = (sum & 0xFFFF) + (sum >> 16);
1661 	ntcph->th_sum = (sum & 0xffff);
1662 
1663 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1664 	    "==> nxge_do_softlso: first mp $%p (mp_chain $%p) "
1665 	    "mss %d pktlen %d l4_len %d (0x%x) "
1666 	    "MBLKL(mp): %d, MBLKL(datamp): %d "
1667 	    "ip_sum 0x%x "
1668 	    "th_sum 0x%x sum 0x%x ) "
1669 	    "dump first ip->tcp %s",
1670 	    nmp, mp_chain,
1671 	    mss,
1672 	    pktlen,
1673 	    l4_len,
1674 	    l4_len,
1675 	    (int)MBLKL(mp), (int)MBLKL(datamp),
1676 	    niph->ip_sum,
1677 	    ntcph->th_sum,
1678 	    sum,
1679 	    nxge_dump_packet((char *)niph, 52)));
1680 
1681 	cmp = nmp;
1682 	while ((nmp = nmp->b_next)->b_next != NULL) {
1683 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1684 		    "==>nxge_do_softlso: middle l4_len %d ", l4_len));
1685 		bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1686 		nmp->b_wptr = nmp->b_rptr + hdrlen;
1687 		niph = (struct ip *)(nmp->b_rptr + ehlen);
1688 		niph->ip_id = htons(++ip_id);
1689 		niph->ip_len = htons(mss + iphlen + tcphlen);
1690 		ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1691 		tcp_seq += mss;
1692 
1693 		ntcph->th_flags &= ~(TH_FIN | TH_PUSH | TH_RST | TH_URG);
1694 
1695 #ifdef __sparc
1696 		tcp_seq_tmp = htonl(tcp_seq);
1697 		bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1698 #else
1699 		ntcph->th_seq = htonl(tcp_seq);
1700 #endif
1701 		DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1702 		DB_CKSUMSTART(nmp) = start_offset;
1703 		DB_CKSUMSTUFF(nmp) = stuff_offset;
1704 
1705 		/* calculate IP checksum and TCP pseudo header checksum */
1706 		niph->ip_sum = 0;
1707 		niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1708 		ntcph->th_sum = (sum & 0xffff);
1709 
1710 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1711 		    "==> nxge_do_softlso: middle ip_sum 0x%x "
1712 		    "th_sum 0x%x "
1713 		    " mp $%p (mp_chain $%p) pktlen %d "
1714 		    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1715 		    niph->ip_sum,
1716 		    ntcph->th_sum,
1717 		    nmp, mp_chain,
1718 		    pktlen, (int)MBLKL(mp), (int)MBLKL(datamp)));
1719 	}
1720 
1721 	/* Last segment */
1722 	/*
1723 	 * Set FIN and/or PSH flags if present only in the last packet.
1724 	 * The ip_len could be different from prior packets.
1725 	 */
1726 	bcopy(cmp->b_rptr, nmp->b_rptr, hdrlen);
1727 	nmp->b_wptr = nmp->b_rptr + hdrlen;
1728 	niph = (struct ip *)(nmp->b_rptr + ehlen);
1729 	niph->ip_id = htons(++ip_id);
1730 	niph->ip_len = htons(msgsize(nmp->b_cont) + iphlen + tcphlen);
1731 	ntcph = (struct tcphdr *)(nmp->b_rptr + ehlen + iphlen);
1732 	tcp_seq += mss;
1733 #ifdef __sparc
1734 	tcp_seq_tmp = htonl(tcp_seq);
1735 	bcopy(&tcp_seq_tmp, (char *)&ntcph->th_seq, 4);
1736 #else
1737 	ntcph->th_seq = htonl(tcp_seq);
1738 #endif
1739 	ntcph->th_flags = (otcph->th_flags & ~TH_URG);
1740 
1741 	DB_CKSUMFLAGS(nmp) = (uint16_t)hckflags;
1742 	DB_CKSUMSTART(nmp) = start_offset;
1743 	DB_CKSUMSTUFF(nmp) = stuff_offset;
1744 
1745 	/* calculate IP checksum and TCP pseudo header checksum */
1746 	niph->ip_sum = 0;
1747 	niph->ip_sum = (uint16_t)nxge_csgen((uint16_t *)niph, iphlen);
1748 
1749 	l4_len = ntohs(niph->ip_len) - iphlen;
1750 	sum = htons(l4_len) + l4cksum;
1751 	sum = (sum & 0xFFFF) + (sum >> 16);
1752 	ntcph->th_sum = (sum & 0xffff);
1753 
1754 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1755 	    "==> nxge_do_softlso: last next "
1756 	    "niph->ip_sum 0x%x "
1757 	    "ntcph->th_sum 0x%x sum 0x%x "
1758 	    "dump last ip->tcp %s "
1759 	    "cmp $%p mp $%p (mp_chain $%p) pktlen %d (0x%x) "
1760 	    "l4_len %d (0x%x) "
1761 	    "MBLKL(mp): %d, MBLKL(datamp): %d ",
1762 	    niph->ip_sum,
1763 	    ntcph->th_sum, sum,
1764 	    nxge_dump_packet((char *)niph, 52),
1765 	    cmp, nmp, mp_chain,
1766 	    pktlen, pktlen,
1767 	    l4_len,
1768 	    l4_len,
1769 	    (int)MBLKL(mp), (int)MBLKL(datamp)));
1770 
1771 cleanup_allocated_msgs:
1772 	if (do_cleanup) {
1773 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1774 		    "==> nxge_do_softlso: "
1775 		    "Failed allocating messages, "
1776 		    "have to clean up and fail!"));
1777 		while (mp_chain != NULL) {
1778 			nmp = mp_chain;
1779 			mp_chain = mp_chain->b_next;
1780 			freemsg(nmp);
1781 		}
1782 	}
1783 	/*
1784 	 * We're done here, so just free the original message and return the
1785 	 * new message chain, that could be NULL if failed, back to the caller.
1786 	 */
1787 	freemsg(mp);
1788 
1789 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1790 	    "<== nxge_do_softlso:mp_chain $%p", mp_chain));
1791 	return (mp_chain);
1792 }
1793 
1794 /*
1795  * Will be called before NIC driver do further operation on the message.
1796  * The input message may include LSO information, if so, go to softlso logic
1797  * to eliminate the oversized LSO packet for the incapable underlying h/w.
1798  * The return could be the same non-LSO message or a message chain for LSO case.
1799  *
1800  * The driver needs to call this function per packet and process the whole chain
1801  * if applied.
1802  */
1803 static mblk_t *
1804 nxge_lso_eliminate(mblk_t *mp)
1805 {
1806 	uint32_t lsoflags;
1807 	uint32_t mss;
1808 
1809 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1810 	    "==>nxge_lso_eliminate:"));
1811 	nxge_lso_info_get(mp, &mss, &lsoflags);
1812 
1813 	if (lsoflags & HW_LSO) {
1814 		mblk_t *nmp;
1815 
1816 		NXGE_DEBUG_MSG((NULL, TX_CTL,
1817 		    "==>nxge_lso_eliminate:"
1818 		    "HW_LSO:mss %d mp $%p",
1819 		    mss, mp));
1820 		if ((nmp = nxge_do_softlso(mp, mss)) != NULL) {
1821 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1822 			    "<== nxge_lso_eliminate: "
1823 			    "LSO: nmp not NULL nmp $%p mss %d mp $%p",
1824 			    nmp, mss, mp));
1825 			return (nmp);
1826 		} else {
1827 			NXGE_DEBUG_MSG((NULL, TX_CTL,
1828 			    "<== nxge_lso_eliminate_ "
1829 			    "LSO: failed nmp NULL nmp $%p mss %d mp $%p",
1830 			    nmp, mss, mp));
1831 			return (NULL);
1832 		}
1833 	}
1834 
1835 	NXGE_DEBUG_MSG((NULL, TX_CTL,
1836 	    "<== nxge_lso_eliminate"));
1837 	return (mp);
1838 }
1839 
1840 static uint32_t
1841 nxge_csgen(uint16_t *adr, int len)
1842 {
1843 	int		i, odd;
1844 	uint32_t	sum = 0;
1845 	uint32_t	c = 0;
1846 
1847 	odd = len % 2;
1848 	for (i = 0; i < (len / 2); i++) {
1849 		sum += (adr[i] & 0xffff);
1850 	}
1851 	if (odd) {
1852 		sum += adr[len / 2] & 0xff00;
1853 	}
1854 	while ((c = ((sum & 0xffff0000) >> 16)) != 0) {
1855 		sum &= 0xffff;
1856 		sum += c;
1857 	}
1858 	return (~sum & 0xffff);
1859 }
1860