xref: /titanic_50/usr/src/uts/common/io/igb/igb_tx.c (revision ac7f5757903d7806e03e59f71c10eec36e0deade)
1 /*
2  * CDDL HEADER START
3  *
4  * Copyright(c) 2007-2009 Intel Corporation. All rights reserved.
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at:
10  *	http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When using or redistributing this file, you may do so under the
15  * License only. No other modification of this header is permitted.
16  *
17  * If applicable, add the following below this CDDL HEADER, with the
18  * fields enclosed by brackets "[]" replaced with your own identifying
19  * information: Portions Copyright [yyyy] [name of copyright owner]
20  *
21  * CDDL HEADER END
22  */
23 
24 /*
25  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms.
27  */
28 
29 #include "igb_sw.h"
30 
31 static boolean_t igb_tx(igb_tx_ring_t *, mblk_t *);
32 static int igb_tx_copy(igb_tx_ring_t *, tx_control_block_t *, mblk_t *,
33     uint32_t, boolean_t);
34 static int igb_tx_bind(igb_tx_ring_t *, tx_control_block_t *, mblk_t *,
35     uint32_t);
36 static int igb_tx_fill_ring(igb_tx_ring_t *, link_list_t *, tx_context_t *,
37     size_t);
38 static void igb_save_desc(tx_control_block_t *, uint64_t, size_t);
39 static tx_control_block_t *igb_get_free_list(igb_tx_ring_t *);
40 static int igb_get_tx_context(mblk_t *, tx_context_t *);
41 static boolean_t igb_check_tx_context(igb_tx_ring_t *, tx_context_t *);
42 static void igb_fill_tx_context(struct e1000_adv_tx_context_desc *,
43     tx_context_t *, uint32_t);
44 
45 #ifndef IGB_DEBUG
46 #pragma inline(igb_save_desc)
47 #pragma inline(igb_get_tx_context)
48 #pragma inline(igb_check_tx_context)
49 #pragma inline(igb_fill_tx_context)
50 #endif
51 
52 mblk_t *
53 igb_tx_ring_send(void *arg, mblk_t *mp)
54 {
55 	igb_tx_ring_t *tx_ring = (igb_tx_ring_t *)arg;
56 
57 	ASSERT(tx_ring != NULL);
58 
59 	if ((tx_ring->igb->igb_state & IGB_SUSPENDED) ||
60 	    (tx_ring->igb->igb_state & IGB_ERROR) ||
61 	    !(tx_ring->igb->igb_state & IGB_STARTED)) {
62 		freemsg(mp);
63 		return (NULL);
64 	}
65 
66 	return ((igb_tx(tx_ring, mp)) ? NULL : mp);
67 }
68 
69 /*
70  * igb_tx - Main transmit processing
71  *
72  * Called from igb_m_tx with an mblk ready to transmit. this
73  * routine sets up the transmit descriptors and sends data to
74  * the wire.
75  *
76  * One mblk can consist of several fragments, each fragment
77  * will be processed with different methods based on the size.
78  * For the fragments with size less than the bcopy threshold,
79  * they will be processed by using bcopy; otherwise, they will
80  * be processed by using DMA binding.
81  *
82  * To process the mblk, a tx control block is got from the
83  * free list. One tx control block contains one tx buffer, which
84  * is used to copy mblk fragments' data; and one tx DMA handle,
85  * which is used to bind a mblk fragment with DMA resource.
86  *
87  * Several small mblk fragments can be copied into one tx control
88  * block's buffer, and then the buffer will be transmitted with
89  * one tx descriptor.
90  *
91  * A large fragment only binds with one tx control block's DMA
92  * handle, and it can span several tx descriptors for transmitting.
93  *
94  * So to transmit a packet (mblk), several tx control blocks can
95  * be used. After the processing, those tx control blocks will
96  * be put to the work list.
97  */
98 static boolean_t
99 igb_tx(igb_tx_ring_t *tx_ring, mblk_t *mp)
100 {
101 	igb_t *igb = tx_ring->igb;
102 	tx_type_t current_flag, next_flag;
103 	uint32_t current_len, next_len;
104 	uint32_t desc_total;
105 	size_t mbsize;
106 	int desc_num;
107 	boolean_t copy_done, eop;
108 	mblk_t *current_mp, *next_mp, *nmp;
109 	tx_control_block_t *tcb;
110 	tx_context_t tx_context, *ctx;
111 	link_list_t pending_list;
112 	mblk_t *hdr_new_mp = NULL;
113 	mblk_t *hdr_previous_mp = NULL;
114 	mblk_t *hdr_current_mp = NULL;
115 	uint32_t hdr_frag_len;
116 	uint32_t hdr_len, len;
117 	uint32_t copy_thresh;
118 
119 	copy_thresh = igb->tx_copy_thresh;
120 
121 	/* Get the mblk size */
122 	mbsize = 0;
123 	for (nmp = mp; nmp != NULL; nmp = nmp->b_cont) {
124 		mbsize += MBLKL(nmp);
125 	}
126 
127 	if (igb->tx_hcksum_enable) {
128 		ctx = &tx_context;
129 		/*
130 		 * Retrieve offloading context information from the mblk
131 		 * that will be used to decide whether/how to fill the
132 		 * context descriptor.
133 		 */
134 		if (igb_get_tx_context(mp, ctx) != TX_CXT_SUCCESS) {
135 			freemsg(mp);
136 			return (B_TRUE);
137 		}
138 
139 		if ((ctx->lso_flag &&
140 		    (mbsize > (ctx->mac_hdr_len + IGB_LSO_MAXLEN))) ||
141 		    (!ctx->lso_flag &&
142 		    (mbsize > (igb->max_frame_size - ETHERFCSL)))) {
143 			freemsg(mp);
144 			IGB_DEBUGLOG_0(igb, "igb_tx: packet oversize");
145 			return (B_TRUE);
146 		}
147 	} else {
148 		ctx = NULL;
149 		if (mbsize > (igb->max_frame_size - ETHERFCSL)) {
150 			freemsg(mp);
151 			IGB_DEBUGLOG_0(igb, "igb_tx: packet oversize");
152 			return (B_TRUE);
153 		}
154 	}
155 
156 	/*
157 	 * Check and recycle tx descriptors.
158 	 * The recycle threshold here should be selected carefully
159 	 */
160 	if (tx_ring->tbd_free < igb->tx_recycle_thresh)
161 		tx_ring->tx_recycle(tx_ring);
162 
163 	/*
164 	 * After the recycling, if the tbd_free is less than the
165 	 * tx_overload_threshold, assert overload, return B_FALSE;
166 	 * and we need to re-schedule the tx again.
167 	 */
168 	if (tx_ring->tbd_free < igb->tx_overload_thresh) {
169 		tx_ring->reschedule = B_TRUE;
170 		IGB_DEBUG_STAT(tx_ring->stat_overload);
171 		return (B_FALSE);
172 	}
173 
174 	/*
175 	 * The software should guarantee LSO packet header(MAC+IP+TCP)
176 	 * to be within one descriptor - this is required by h/w.
177 	 * Here will reallocate and refill the header if
178 	 * the headers(MAC+IP+TCP) is physical memory non-contiguous.
179 	 */
180 	if (ctx && ctx->lso_flag) {
181 		hdr_len = ctx->mac_hdr_len + ctx->ip_hdr_len + ctx->l4_hdr_len;
182 		len = MBLKL(mp);
183 		hdr_current_mp = mp;
184 		while (len < hdr_len) {
185 			hdr_previous_mp = hdr_current_mp;
186 			hdr_current_mp = hdr_current_mp->b_cont;
187 			len += MBLKL(hdr_current_mp);
188 		}
189 		/*
190 		 * If the header and the payload are in different mblks,
191 		 * we simply force the header to be copied into pre-allocated
192 		 * page-aligned buffer.
193 		 */
194 		if (len == hdr_len)
195 			goto adjust_threshold;
196 
197 		hdr_frag_len = hdr_len - (len - MBLKL(hdr_current_mp));
198 		/*
199 		 * There are two cases we will reallocate
200 		 * a mblk for the last header fragment.
201 		 * 1. the header is in multiple mblks and
202 		 *    the last fragment shares the same mblk
203 		 *    with the payload
204 		 * 2. the header is in a single mblk shared
205 		 *    with the payload but the header crosses
206 		 *    a page.
207 		 */
208 		if ((hdr_current_mp != mp) ||
209 		    (P2NPHASE((uintptr_t)hdr_current_mp->b_rptr, igb->page_size)
210 		    < hdr_len)) {
211 			/*
212 			 * reallocate the mblk for the last header fragment,
213 			 * expect it to be copied into pre-allocated
214 			 * page-aligned buffer
215 			 */
216 			hdr_new_mp = allocb(hdr_frag_len, NULL);
217 			if (!hdr_new_mp) {
218 				return (B_FALSE);
219 			}
220 
221 			/* link the new header fragment with the other parts */
222 			bcopy(hdr_current_mp->b_rptr,
223 			    hdr_new_mp->b_rptr, hdr_frag_len);
224 			hdr_new_mp->b_wptr = hdr_new_mp->b_rptr + hdr_frag_len;
225 			hdr_new_mp->b_cont = hdr_current_mp;
226 			if (hdr_previous_mp)
227 				hdr_previous_mp->b_cont = hdr_new_mp;
228 			else
229 				mp = hdr_new_mp;
230 			hdr_current_mp->b_rptr += hdr_frag_len;
231 		}
232 adjust_threshold:
233 		/*
234 		 * adjust the bcopy threshhold to guarantee
235 		 * the header to use bcopy way
236 		 */
237 		if (copy_thresh < hdr_len)
238 			copy_thresh = hdr_len;
239 	}
240 
241 	/*
242 	 * The pending_list is a linked list that is used to save
243 	 * the tx control blocks that have packet data processed
244 	 * but have not put the data to the tx descriptor ring.
245 	 * It is used to reduce the lock contention of the tx_lock.
246 	 */
247 	LINK_LIST_INIT(&pending_list);
248 	desc_num = 0;
249 	desc_total = 0;
250 
251 	current_mp = mp;
252 	current_len = MBLKL(current_mp);
253 	/*
254 	 * Decide which method to use for the first fragment
255 	 */
256 	current_flag = (current_len <= copy_thresh) ?
257 	    USE_COPY : USE_DMA;
258 	/*
259 	 * If the mblk includes several contiguous small fragments,
260 	 * they may be copied into one buffer. This flag is used to
261 	 * indicate whether there are pending fragments that need to
262 	 * be copied to the current tx buffer.
263 	 *
264 	 * If this flag is B_TRUE, it indicates that a new tx control
265 	 * block is needed to process the next fragment using either
266 	 * copy or DMA binding.
267 	 *
268 	 * Otherwise, it indicates that the next fragment will be
269 	 * copied to the current tx buffer that is maintained by the
270 	 * current tx control block. No new tx control block is needed.
271 	 */
272 	copy_done = B_TRUE;
273 	while (current_mp) {
274 		next_mp = current_mp->b_cont;
275 		eop = (next_mp == NULL); /* Last fragment of the packet? */
276 		next_len = eop ? 0: MBLKL(next_mp);
277 
278 		/*
279 		 * When the current fragment is an empty fragment, if
280 		 * the next fragment will still be copied to the current
281 		 * tx buffer, we cannot skip this fragment here. Because
282 		 * the copy processing is pending for completion. We have
283 		 * to process this empty fragment in the tx_copy routine.
284 		 *
285 		 * If the copy processing is completed or a DMA binding
286 		 * processing is just completed, we can just skip this
287 		 * empty fragment.
288 		 */
289 		if ((current_len == 0) && (copy_done)) {
290 			current_mp = next_mp;
291 			current_len = next_len;
292 			current_flag = (current_len <= copy_thresh) ?
293 			    USE_COPY : USE_DMA;
294 			continue;
295 		}
296 
297 		if (copy_done) {
298 			/*
299 			 * Get a new tx control block from the free list
300 			 */
301 			tcb = igb_get_free_list(tx_ring);
302 
303 			if (tcb == NULL) {
304 				IGB_DEBUG_STAT(tx_ring->stat_fail_no_tcb);
305 				goto tx_failure;
306 			}
307 
308 			/*
309 			 * Push the tx control block to the pending list
310 			 * to avoid using lock too early
311 			 */
312 			LIST_PUSH_TAIL(&pending_list, &tcb->link);
313 		}
314 
315 		if (current_flag == USE_COPY) {
316 			/*
317 			 * Check whether to use bcopy or DMA binding to process
318 			 * the next fragment, and if using bcopy, whether we
319 			 * need to continue copying the next fragment into the
320 			 * current tx buffer.
321 			 */
322 			ASSERT((tcb->tx_buf.len + current_len) <=
323 			    tcb->tx_buf.size);
324 
325 			if (eop) {
326 				/*
327 				 * This is the last fragment of the packet, so
328 				 * the copy processing will be completed with
329 				 * this fragment.
330 				 */
331 				next_flag = USE_NONE;
332 				copy_done = B_TRUE;
333 			} else if ((tcb->tx_buf.len + current_len + next_len) >
334 			    tcb->tx_buf.size) {
335 				/*
336 				 * If the next fragment is too large to be
337 				 * copied to the current tx buffer, we need
338 				 * to complete the current copy processing.
339 				 */
340 				next_flag = (next_len > copy_thresh) ?
341 				    USE_DMA: USE_COPY;
342 				copy_done = B_TRUE;
343 			} else if (next_len > copy_thresh) {
344 				/*
345 				 * The next fragment needs to be processed with
346 				 * DMA binding. So the copy prcessing will be
347 				 * completed with the current fragment.
348 				 */
349 				next_flag = USE_DMA;
350 				copy_done = B_TRUE;
351 			} else {
352 				/*
353 				 * Continue to copy the next fragment to the
354 				 * current tx buffer.
355 				 */
356 				next_flag = USE_COPY;
357 				copy_done = B_FALSE;
358 			}
359 
360 			desc_num = igb_tx_copy(tx_ring, tcb, current_mp,
361 			    current_len, copy_done);
362 		} else {
363 			/*
364 			 * Check whether to use bcopy or DMA binding to process
365 			 * the next fragment.
366 			 */
367 			next_flag = (next_len > copy_thresh) ?
368 			    USE_DMA: USE_COPY;
369 			ASSERT(copy_done == B_TRUE);
370 
371 			desc_num = igb_tx_bind(tx_ring, tcb, current_mp,
372 			    current_len);
373 		}
374 
375 		if (desc_num > 0)
376 			desc_total += desc_num;
377 		else if (desc_num < 0)
378 			goto tx_failure;
379 
380 		current_mp = next_mp;
381 		current_len = next_len;
382 		current_flag = next_flag;
383 	}
384 
385 	/*
386 	 * Attach the mblk to the last tx control block
387 	 */
388 	ASSERT(tcb);
389 	ASSERT(tcb->mp == NULL);
390 	tcb->mp = mp;
391 
392 	/*
393 	 * Before fill the tx descriptor ring with the data, we need to
394 	 * ensure there are adequate free descriptors for transmit
395 	 * (including one context descriptor).
396 	 */
397 	if (tx_ring->tbd_free < (desc_total + 1)) {
398 		tx_ring->tx_recycle(tx_ring);
399 	}
400 
401 	mutex_enter(&tx_ring->tx_lock);
402 
403 	/*
404 	 * If the number of free tx descriptors is not enough for transmit
405 	 * then return failure.
406 	 *
407 	 * Note: we must put this check under the mutex protection to
408 	 * ensure the correctness when multiple threads access it in
409 	 * parallel.
410 	 */
411 	if (tx_ring->tbd_free < (desc_total + 1)) {
412 		IGB_DEBUG_STAT(tx_ring->stat_fail_no_tbd);
413 		mutex_exit(&tx_ring->tx_lock);
414 		goto tx_failure;
415 	}
416 
417 	desc_num = igb_tx_fill_ring(tx_ring, &pending_list, ctx, mbsize);
418 
419 	ASSERT((desc_num == desc_total) || (desc_num == (desc_total + 1)));
420 
421 	mutex_exit(&tx_ring->tx_lock);
422 
423 	return (B_TRUE);
424 
425 tx_failure:
426 	/*
427 	 * If new mblk has been allocted for the last header
428 	 * fragment of a LSO packet, we should restore the
429 	 * modified mp.
430 	 */
431 	if (hdr_new_mp) {
432 		hdr_new_mp->b_cont = NULL;
433 		freeb(hdr_new_mp);
434 		hdr_current_mp->b_rptr -= hdr_frag_len;
435 		if (hdr_previous_mp)
436 			hdr_previous_mp->b_cont = hdr_current_mp;
437 		else
438 			mp = hdr_current_mp;
439 	}
440 
441 	/*
442 	 * Discard the mblk and free the used resources
443 	 */
444 	tcb = (tx_control_block_t *)LIST_GET_HEAD(&pending_list);
445 	while (tcb) {
446 		tcb->mp = NULL;
447 
448 		igb_free_tcb(tcb);
449 
450 		tcb = (tx_control_block_t *)
451 		    LIST_GET_NEXT(&pending_list, &tcb->link);
452 	}
453 
454 	/*
455 	 * Return the tx control blocks in the pending list to the free list.
456 	 */
457 	igb_put_free_list(tx_ring, &pending_list);
458 
459 	/* Transmit failed, do not drop the mblk, rechedule the transmit */
460 	tx_ring->reschedule = B_TRUE;
461 
462 	return (B_FALSE);
463 }
464 
465 /*
466  * igb_tx_copy
467  *
468  * Copy the mblk fragment to the pre-allocated tx buffer
469  */
470 static int
471 igb_tx_copy(igb_tx_ring_t *tx_ring, tx_control_block_t *tcb, mblk_t *mp,
472     uint32_t len, boolean_t copy_done)
473 {
474 	dma_buffer_t *tx_buf;
475 	uint32_t desc_num;
476 	_NOTE(ARGUNUSED(tx_ring));
477 
478 	tx_buf = &tcb->tx_buf;
479 
480 	/*
481 	 * Copy the packet data of the mblk fragment into the
482 	 * pre-allocated tx buffer, which is maintained by the
483 	 * tx control block.
484 	 *
485 	 * Several mblk fragments can be copied into one tx buffer.
486 	 * The destination address of the current copied fragment in
487 	 * the tx buffer is next to the end of the previous copied
488 	 * fragment.
489 	 */
490 	if (len > 0) {
491 		bcopy(mp->b_rptr, tx_buf->address + tx_buf->len, len);
492 
493 		tx_buf->len += len;
494 		tcb->frag_num++;
495 	}
496 
497 	desc_num = 0;
498 
499 	/*
500 	 * If it is the last fragment copied to the current tx buffer,
501 	 * in other words, if there's no remaining fragment or the remaining
502 	 * fragment requires a new tx control block to process, we need to
503 	 * complete the current copy processing by syncing up the current
504 	 * DMA buffer and saving the descriptor data.
505 	 */
506 	if (copy_done) {
507 		/*
508 		 * Sync the DMA buffer of the packet data
509 		 */
510 		DMA_SYNC(tx_buf, DDI_DMA_SYNC_FORDEV);
511 
512 		tcb->tx_type = USE_COPY;
513 
514 		/*
515 		 * Save the address and length to the private data structure
516 		 * of the tx control block, which will be used to fill the
517 		 * tx descriptor ring after all the fragments are processed.
518 		 */
519 		igb_save_desc(tcb, tx_buf->dma_address, tx_buf->len);
520 		desc_num++;
521 	}
522 
523 	return (desc_num);
524 }
525 
526 /*
527  * igb_tx_bind
528  *
529  * Bind the mblk fragment with DMA
530  */
531 static int
532 igb_tx_bind(igb_tx_ring_t *tx_ring, tx_control_block_t *tcb, mblk_t *mp,
533     uint32_t len)
534 {
535 	int status, i;
536 	ddi_dma_cookie_t dma_cookie;
537 	uint_t ncookies;
538 	int desc_num;
539 
540 	/*
541 	 * Use DMA binding to process the mblk fragment
542 	 */
543 	status = ddi_dma_addr_bind_handle(tcb->tx_dma_handle, NULL,
544 	    (caddr_t)mp->b_rptr, len,
545 	    DDI_DMA_WRITE | DDI_DMA_STREAMING, DDI_DMA_DONTWAIT,
546 	    0, &dma_cookie, &ncookies);
547 
548 	if (status != DDI_DMA_MAPPED) {
549 		IGB_DEBUG_STAT(tx_ring->stat_fail_dma_bind);
550 		return (-1);
551 	}
552 
553 	tcb->frag_num++;
554 	tcb->tx_type = USE_DMA;
555 	/*
556 	 * Each fragment can span several cookies. One cookie will have
557 	 * one tx descriptor to transmit.
558 	 */
559 	desc_num = 0;
560 	for (i = ncookies; i > 0; i--) {
561 		/*
562 		 * Save the address and length to the private data structure
563 		 * of the tx control block, which will be used to fill the
564 		 * tx descriptor ring after all the fragments are processed.
565 		 */
566 		igb_save_desc(tcb,
567 		    dma_cookie.dmac_laddress,
568 		    dma_cookie.dmac_size);
569 
570 		desc_num++;
571 
572 		if (i > 1)
573 			ddi_dma_nextcookie(tcb->tx_dma_handle, &dma_cookie);
574 	}
575 
576 	return (desc_num);
577 }
578 
579 /*
580  * igb_get_tx_context
581  *
582  * Get the tx context information from the mblk
583  */
584 static int
585 igb_get_tx_context(mblk_t *mp, tx_context_t *ctx)
586 {
587 	uint32_t start;
588 	uint32_t flags;
589 	uint32_t lso_flag;
590 	uint32_t mss;
591 	uint32_t len;
592 	uint32_t size;
593 	uint32_t offset;
594 	unsigned char *pos;
595 	ushort_t etype;
596 	uint32_t mac_hdr_len;
597 	uint32_t l4_proto;
598 	uint32_t l4_hdr_len;
599 
600 	ASSERT(mp != NULL);
601 
602 	hcksum_retrieve(mp, NULL, NULL, &start, NULL, NULL, NULL, &flags);
603 	bzero(ctx, sizeof (tx_context_t));
604 
605 	ctx->hcksum_flags = flags;
606 
607 	if (flags == 0)
608 		return (TX_CXT_SUCCESS);
609 
610 	lso_info_get(mp, &mss, &lso_flag);
611 	ctx->mss = mss;
612 	ctx->lso_flag = (lso_flag == HW_LSO);
613 
614 	/*
615 	 * LSO relies on tx h/w checksum, so here the packet will be
616 	 * dropped if the h/w checksum flags are not set.
617 	 */
618 	if (ctx->lso_flag) {
619 		if (!((ctx->hcksum_flags & HCK_PARTIALCKSUM) &&
620 		    (ctx->hcksum_flags & HCK_IPV4_HDRCKSUM))) {
621 			IGB_DEBUGLOG_0(NULL, "igb_tx: h/w "
622 			    "checksum flags are not set for LSO");
623 			return (TX_CXT_E_LSO_CSUM);
624 		}
625 	}
626 
627 	etype = 0;
628 	mac_hdr_len = 0;
629 	l4_proto = 0;
630 
631 	/*
632 	 * Firstly get the position of the ether_type/ether_tpid.
633 	 * Here we don't assume the ether (VLAN) header is fully included
634 	 * in one mblk fragment, so we go thourgh the fragments to parse
635 	 * the ether type.
636 	 */
637 	size = len = MBLKL(mp);
638 	offset = offsetof(struct ether_header, ether_type);
639 	while (size <= offset) {
640 		mp = mp->b_cont;
641 		ASSERT(mp != NULL);
642 		len = MBLKL(mp);
643 		size += len;
644 	}
645 	pos = mp->b_rptr + offset + len - size;
646 
647 	etype = ntohs(*(ushort_t *)(uintptr_t)pos);
648 	if (etype == ETHERTYPE_VLAN) {
649 		/*
650 		 * Get the position of the ether_type in VLAN header
651 		 */
652 		offset = offsetof(struct ether_vlan_header, ether_type);
653 		while (size <= offset) {
654 			mp = mp->b_cont;
655 			ASSERT(mp != NULL);
656 			len = MBLKL(mp);
657 			size += len;
658 		}
659 		pos = mp->b_rptr + offset + len - size;
660 
661 		etype = ntohs(*(ushort_t *)(uintptr_t)pos);
662 		mac_hdr_len = sizeof (struct ether_vlan_header);
663 	} else {
664 		mac_hdr_len = sizeof (struct ether_header);
665 	}
666 
667 	/*
668 	 * Here we assume the IP(V6) header is fully included in one
669 	 * mblk fragment.
670 	 */
671 	switch (etype) {
672 	case ETHERTYPE_IP:
673 		offset = mac_hdr_len;
674 		while (size <= offset) {
675 			mp = mp->b_cont;
676 			ASSERT(mp != NULL);
677 			len = MBLKL(mp);
678 			size += len;
679 		}
680 		pos = mp->b_rptr + offset + len - size;
681 
682 		if (ctx->lso_flag) {
683 			*((uint16_t *)(uintptr_t)(pos + offsetof(ipha_t,
684 			    ipha_length))) = 0;
685 
686 			/*
687 			 * To utilize igb LSO, here need to fill
688 			 * the tcp checksum field of the packet with the
689 			 * following pseudo-header checksum:
690 			 * (ip_source_addr, ip_destination_addr, l4_proto)
691 			 * and also need to fill the ip header checksum
692 			 * with zero. Currently the tcp/ip stack has done
693 			 * these.
694 			 */
695 		}
696 
697 		l4_proto = *(uint8_t *)(pos + offsetof(ipha_t, ipha_protocol));
698 		break;
699 	case ETHERTYPE_IPV6:
700 		offset = offsetof(ip6_t, ip6_nxt) + mac_hdr_len;
701 		while (size <= offset) {
702 			mp = mp->b_cont;
703 			ASSERT(mp != NULL);
704 			len = MBLKL(mp);
705 			size += len;
706 		}
707 		pos = mp->b_rptr + offset + len - size;
708 
709 		l4_proto = *(uint8_t *)pos;
710 		break;
711 	default:
712 		/* Unrecoverable error */
713 		IGB_DEBUGLOG_0(NULL, "Ethernet type field error with "
714 		    "tx hcksum flag set");
715 		return (TX_CXT_E_ETHER_TYPE);
716 	}
717 
718 	if (ctx->lso_flag) {
719 		offset = mac_hdr_len + start;
720 		while (size <= offset) {
721 			mp = mp->b_cont;
722 			ASSERT(mp != NULL);
723 			len = MBLKL(mp);
724 			size += len;
725 		}
726 		pos = mp->b_rptr + offset + len - size;
727 
728 		l4_hdr_len = TCP_HDR_LENGTH((tcph_t *)pos);
729 	} else {
730 		/*
731 		 * l4 header length is only required for LSO
732 		 */
733 		l4_hdr_len = 0;
734 	}
735 
736 	ctx->mac_hdr_len = mac_hdr_len;
737 	ctx->ip_hdr_len = start;
738 	ctx->l4_proto = l4_proto;
739 	ctx->l4_hdr_len = l4_hdr_len;
740 
741 	return (TX_CXT_SUCCESS);
742 }
743 
744 /*
745  * igb_check_tx_context
746  *
747  * Check if a new context descriptor is needed
748  */
749 static boolean_t
750 igb_check_tx_context(igb_tx_ring_t *tx_ring, tx_context_t *ctx)
751 {
752 	tx_context_t *last;
753 
754 	if (ctx == NULL)
755 		return (B_FALSE);
756 
757 	/*
758 	 * Compare the context data retrieved from the mblk and the
759 	 * stored context data of the last context descriptor. The data
760 	 * need to be checked are:
761 	 *	hcksum_flags
762 	 *	l4_proto
763 	 *	mss (only check for LSO)
764 	 *	l4_hdr_len (only check for LSO)
765 	 *	ip_hdr_len
766 	 *	mac_hdr_len
767 	 * Either one of the above data is changed, a new context descriptor
768 	 * will be needed.
769 	 */
770 	last = &tx_ring->tx_context;
771 
772 	if (ctx->hcksum_flags != 0) {
773 		if ((ctx->hcksum_flags != last->hcksum_flags) ||
774 		    (ctx->l4_proto != last->l4_proto) ||
775 		    (ctx->lso_flag && ((ctx->mss != last->mss) ||
776 		    (ctx->l4_hdr_len != last->l4_hdr_len))) ||
777 		    (ctx->ip_hdr_len != last->ip_hdr_len) ||
778 		    (ctx->mac_hdr_len != last->mac_hdr_len)) {
779 			return (B_TRUE);
780 		}
781 	}
782 
783 	return (B_FALSE);
784 }
785 
786 /*
787  * igb_fill_tx_context
788  *
789  * Fill the context descriptor with hardware checksum informations
790  */
791 static void
792 igb_fill_tx_context(struct e1000_adv_tx_context_desc *ctx_tbd,
793     tx_context_t *ctx, uint32_t ring_index)
794 {
795 	/*
796 	 * Fill the context descriptor with the checksum
797 	 * context information we've got
798 	 */
799 	ctx_tbd->vlan_macip_lens = ctx->ip_hdr_len;
800 	ctx_tbd->vlan_macip_lens |= ctx->mac_hdr_len <<
801 	    E1000_ADVTXD_MACLEN_SHIFT;
802 
803 	ctx_tbd->type_tucmd_mlhl =
804 	    E1000_ADVTXD_DCMD_DEXT | E1000_ADVTXD_DTYP_CTXT;
805 
806 	if (ctx->hcksum_flags & HCK_IPV4_HDRCKSUM)
807 		ctx_tbd->type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_IPV4;
808 
809 	if (ctx->hcksum_flags & HCK_PARTIALCKSUM) {
810 		switch (ctx->l4_proto) {
811 		case IPPROTO_TCP:
812 			ctx_tbd->type_tucmd_mlhl |= E1000_ADVTXD_TUCMD_L4T_TCP;
813 			break;
814 		case IPPROTO_UDP:
815 			/*
816 			 * We don't have to explicitly set:
817 			 *	ctx_tbd->type_tucmd_mlhl |=
818 			 *	    E1000_ADVTXD_TUCMD_L4T_UDP;
819 			 * Because E1000_ADVTXD_TUCMD_L4T_UDP == 0b
820 			 */
821 			break;
822 		default:
823 			/* Unrecoverable error */
824 			IGB_DEBUGLOG_0(NULL, "L4 type error with tx hcksum");
825 			break;
826 		}
827 	}
828 
829 	ctx_tbd->seqnum_seed = 0;
830 	ctx_tbd->mss_l4len_idx = ring_index << 4;
831 	if (ctx->lso_flag) {
832 		ctx_tbd->mss_l4len_idx |=
833 		    (ctx->l4_hdr_len << E1000_ADVTXD_L4LEN_SHIFT) |
834 		    (ctx->mss << E1000_ADVTXD_MSS_SHIFT);
835 	}
836 }
837 
838 /*
839  * igb_tx_fill_ring
840  *
841  * Fill the tx descriptor ring with the data
842  */
843 static int
844 igb_tx_fill_ring(igb_tx_ring_t *tx_ring, link_list_t *pending_list,
845     tx_context_t *ctx, size_t mbsize)
846 {
847 	struct e1000_hw *hw = &tx_ring->igb->hw;
848 	boolean_t load_context;
849 	uint32_t index, tcb_index, desc_num;
850 	union e1000_adv_tx_desc *tbd, *first_tbd;
851 	tx_control_block_t *tcb, *first_tcb;
852 	uint32_t hcksum_flags;
853 	int i;
854 	igb_t *igb = tx_ring->igb;
855 
856 	ASSERT(mutex_owned(&tx_ring->tx_lock));
857 
858 	tbd = NULL;
859 	first_tbd = NULL;
860 	first_tcb = NULL;
861 	desc_num = 0;
862 	hcksum_flags = 0;
863 	load_context = B_FALSE;
864 
865 	/*
866 	 * Get the index of the first tx descriptor that will be filled,
867 	 * and the index of the first work list item that will be attached
868 	 * with the first used tx control block in the pending list.
869 	 * Note: the two indexes are the same.
870 	 */
871 	index = tx_ring->tbd_tail;
872 	tcb_index = tx_ring->tbd_tail;
873 
874 	if (ctx != NULL) {
875 		hcksum_flags = ctx->hcksum_flags;
876 
877 		/*
878 		 * Check if a new context descriptor is needed for this packet
879 		 */
880 		load_context = igb_check_tx_context(tx_ring, ctx);
881 		if (load_context) {
882 			first_tcb = (tx_control_block_t *)
883 			    LIST_GET_HEAD(pending_list);
884 			tbd = &tx_ring->tbd_ring[index];
885 
886 			/*
887 			 * Fill the context descriptor with the
888 			 * hardware checksum offload informations.
889 			 */
890 			igb_fill_tx_context(
891 			    (struct e1000_adv_tx_context_desc *)tbd,
892 			    ctx, tx_ring->index);
893 
894 			index = NEXT_INDEX(index, 1, tx_ring->ring_size);
895 			desc_num++;
896 
897 			/*
898 			 * Store the checksum context data if
899 			 * a new context descriptor is added
900 			 */
901 			tx_ring->tx_context = *ctx;
902 		}
903 	}
904 
905 	first_tbd = &tx_ring->tbd_ring[index];
906 
907 	/*
908 	 * Fill tx data descriptors with the data saved in the pending list.
909 	 * The tx control blocks in the pending list are added to the work list
910 	 * at the same time.
911 	 *
912 	 * The work list is strictly 1:1 corresponding to the descriptor ring.
913 	 * One item of the work list corresponds to one tx descriptor. Because
914 	 * one tx control block can span multiple tx descriptors, the tx
915 	 * control block will be added to the first work list item that
916 	 * corresponds to the first tx descriptor generated from that tx
917 	 * control block.
918 	 */
919 	tcb = (tx_control_block_t *)LIST_POP_HEAD(pending_list);
920 	while (tcb != NULL) {
921 
922 		for (i = 0; i < tcb->desc_num; i++) {
923 			tbd = &tx_ring->tbd_ring[index];
924 
925 			tbd->read.buffer_addr = tcb->desc[i].address;
926 			tbd->read.cmd_type_len = tcb->desc[i].length;
927 
928 			tbd->read.cmd_type_len |= E1000_ADVTXD_DCMD_RS |
929 			    E1000_ADVTXD_DCMD_DEXT | E1000_ADVTXD_DTYP_DATA |
930 			    E1000_ADVTXD_DCMD_IFCS;
931 
932 			tbd->read.olinfo_status = 0;
933 
934 			index = NEXT_INDEX(index, 1, tx_ring->ring_size);
935 			desc_num++;
936 		}
937 
938 		if (first_tcb != NULL) {
939 			/*
940 			 * Count the checksum context descriptor for
941 			 * the first tx control block.
942 			 */
943 			first_tcb->desc_num++;
944 			first_tcb = NULL;
945 		}
946 
947 		/*
948 		 * Add the tx control block to the work list
949 		 */
950 		ASSERT(tx_ring->work_list[tcb_index] == NULL);
951 		tx_ring->work_list[tcb_index] = tcb;
952 
953 		tcb_index = index;
954 		tcb = (tx_control_block_t *)LIST_POP_HEAD(pending_list);
955 	}
956 
957 	/*
958 	 * The Insert Ethernet CRC (IFCS) bit and the checksum fields are only
959 	 * valid in the first descriptor of the packet.
960 	 * 82576 also requires the payload length setting even without LSO
961 	 */
962 	ASSERT(first_tbd != NULL);
963 	first_tbd->read.cmd_type_len |= E1000_ADVTXD_DCMD_IFCS;
964 	if (ctx != NULL && ctx->lso_flag) {
965 		first_tbd->read.cmd_type_len |= E1000_ADVTXD_DCMD_TSE;
966 		first_tbd->read.olinfo_status |=
967 		    (mbsize - ctx->mac_hdr_len - ctx->ip_hdr_len
968 		    - ctx->l4_hdr_len) << E1000_ADVTXD_PAYLEN_SHIFT;
969 	} else {
970 		if (hw->mac.type >= e1000_82576) {
971 			first_tbd->read.olinfo_status |=
972 			    (mbsize << E1000_ADVTXD_PAYLEN_SHIFT);
973 		}
974 	}
975 
976 	/* Set hardware checksum bits */
977 	if (hcksum_flags != 0) {
978 		if (hcksum_flags & HCK_IPV4_HDRCKSUM)
979 			first_tbd->read.olinfo_status |=
980 			    E1000_TXD_POPTS_IXSM << 8;
981 		if (hcksum_flags & HCK_PARTIALCKSUM)
982 			first_tbd->read.olinfo_status |=
983 			    E1000_TXD_POPTS_TXSM << 8;
984 		first_tbd->read.olinfo_status |= tx_ring->index << 4;
985 	}
986 
987 	/*
988 	 * The last descriptor of packet needs End Of Packet (EOP),
989 	 * and Report Status (RS) bits set
990 	 */
991 	ASSERT(tbd != NULL);
992 	tbd->read.cmd_type_len |=
993 	    E1000_ADVTXD_DCMD_EOP | E1000_ADVTXD_DCMD_RS;
994 
995 	IGB_DEBUG_STAT(tx_ring->stat_pkt_cnt);
996 
997 	/*
998 	 * Sync the DMA buffer of the tx descriptor ring
999 	 */
1000 	DMA_SYNC(&tx_ring->tbd_area, DDI_DMA_SYNC_FORDEV);
1001 
1002 	/*
1003 	 * Update the number of the free tx descriptors.
1004 	 * The mutual exclusion between the transmission and the recycling
1005 	 * (for the tx descriptor ring and the work list) is implemented
1006 	 * with the atomic operation on the number of the free tx descriptors.
1007 	 *
1008 	 * Note: we should always decrement the counter tbd_free before
1009 	 * advancing the hardware TDT pointer to avoid the race condition -
1010 	 * before the counter tbd_free is decremented, the transmit of the
1011 	 * tx descriptors has done and the counter tbd_free is increased by
1012 	 * the tx recycling.
1013 	 */
1014 	i = igb_atomic_reserve(&tx_ring->tbd_free, desc_num);
1015 	ASSERT(i >= 0);
1016 
1017 	tx_ring->tbd_tail = index;
1018 
1019 	/*
1020 	 * Advance the hardware TDT pointer of the tx descriptor ring
1021 	 */
1022 	E1000_WRITE_REG(hw, E1000_TDT(tx_ring->index), index);
1023 
1024 	if (igb_check_acc_handle(igb->osdep.reg_handle) != DDI_FM_OK) {
1025 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
1026 		atomic_or_32(&igb->igb_state, IGB_ERROR);
1027 	}
1028 
1029 	return (desc_num);
1030 }
1031 
1032 /*
1033  * igb_save_desc
1034  *
1035  * Save the address/length pair to the private array
1036  * of the tx control block. The address/length pairs
1037  * will be filled into the tx descriptor ring later.
1038  */
1039 static void
1040 igb_save_desc(tx_control_block_t *tcb, uint64_t address, size_t length)
1041 {
1042 	sw_desc_t *desc;
1043 
1044 	desc = &tcb->desc[tcb->desc_num];
1045 	desc->address = address;
1046 	desc->length = length;
1047 
1048 	tcb->desc_num++;
1049 }
1050 
1051 /*
1052  * igb_tx_recycle_legacy
1053  *
1054  * Recycle the tx descriptors and tx control blocks.
1055  *
1056  * The work list is traversed to check if the corresponding
1057  * tx descriptors have been transmitted. If so, the resources
1058  * bound to the tx control blocks will be freed, and those
1059  * tx control blocks will be returned to the free list.
1060  */
1061 uint32_t
1062 igb_tx_recycle_legacy(igb_tx_ring_t *tx_ring)
1063 {
1064 	uint32_t index, last_index;
1065 	int desc_num;
1066 	boolean_t desc_done;
1067 	tx_control_block_t *tcb;
1068 	link_list_t pending_list;
1069 	igb_t *igb = tx_ring->igb;
1070 
1071 	/*
1072 	 * The mutex_tryenter() is used to avoid unnecessary
1073 	 * lock contention.
1074 	 */
1075 	if (mutex_tryenter(&tx_ring->recycle_lock) == 0)
1076 		return (0);
1077 
1078 	ASSERT(tx_ring->tbd_free <= tx_ring->ring_size);
1079 
1080 	if (tx_ring->tbd_free == tx_ring->ring_size) {
1081 		tx_ring->recycle_fail = 0;
1082 		tx_ring->stall_watchdog = 0;
1083 		mutex_exit(&tx_ring->recycle_lock);
1084 		return (0);
1085 	}
1086 
1087 	/*
1088 	 * Sync the DMA buffer of the tx descriptor ring
1089 	 */
1090 	DMA_SYNC(&tx_ring->tbd_area, DDI_DMA_SYNC_FORKERNEL);
1091 
1092 	if (igb_check_dma_handle(
1093 	    tx_ring->tbd_area.dma_handle) != DDI_FM_OK) {
1094 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
1095 		atomic_or_32(&igb->igb_state, IGB_ERROR);
1096 		return (0);
1097 	}
1098 
1099 	LINK_LIST_INIT(&pending_list);
1100 	desc_num = 0;
1101 	index = tx_ring->tbd_head;	/* Index of next tbd/tcb to recycle */
1102 
1103 	tcb = tx_ring->work_list[index];
1104 	ASSERT(tcb != NULL);
1105 
1106 	desc_done = B_TRUE;
1107 	while (desc_done && (tcb != NULL)) {
1108 
1109 		/*
1110 		 * Get the last tx descriptor of the tx control block.
1111 		 * If the last tx descriptor is done, it is done with
1112 		 * all the tx descriptors of the tx control block.
1113 		 * Then the tx control block and all the corresponding
1114 		 * tx descriptors can be recycled.
1115 		 */
1116 		last_index = NEXT_INDEX(index, tcb->desc_num - 1,
1117 		    tx_ring->ring_size);
1118 
1119 		/*
1120 		 * Check if the Descriptor Done bit is set
1121 		 */
1122 		desc_done = tx_ring->tbd_ring[last_index].wb.status &
1123 		    E1000_TXD_STAT_DD;
1124 		if (desc_done) {
1125 			/*
1126 			 * Strip off the tx control block from the work list,
1127 			 * and add it to the pending list.
1128 			 */
1129 			tx_ring->work_list[index] = NULL;
1130 			LIST_PUSH_TAIL(&pending_list, &tcb->link);
1131 
1132 			/*
1133 			 * Count the total number of the tx descriptors recycled
1134 			 */
1135 			desc_num += tcb->desc_num;
1136 
1137 			/*
1138 			 * Advance the index of the tx descriptor ring
1139 			 */
1140 			index = NEXT_INDEX(last_index, 1, tx_ring->ring_size);
1141 
1142 			tcb = tx_ring->work_list[index];
1143 		}
1144 	}
1145 
1146 	/*
1147 	 * If no tx descriptors are recycled, no need to do more processing
1148 	 */
1149 	if (desc_num == 0) {
1150 		tx_ring->recycle_fail++;
1151 		mutex_exit(&tx_ring->recycle_lock);
1152 		return (0);
1153 	}
1154 
1155 	tx_ring->recycle_fail = 0;
1156 	tx_ring->stall_watchdog = 0;
1157 
1158 	/*
1159 	 * Update the head index of the tx descriptor ring
1160 	 */
1161 	tx_ring->tbd_head = index;
1162 
1163 	/*
1164 	 * Update the number of the free tx descriptors with atomic operations
1165 	 */
1166 	atomic_add_32(&tx_ring->tbd_free, desc_num);
1167 
1168 	mutex_exit(&tx_ring->recycle_lock);
1169 
1170 	/*
1171 	 * Free the resources used by the tx control blocks
1172 	 * in the pending list
1173 	 */
1174 	tcb = (tx_control_block_t *)LIST_GET_HEAD(&pending_list);
1175 	while (tcb != NULL) {
1176 		/*
1177 		 * Release the resources occupied by the tx control block
1178 		 */
1179 		igb_free_tcb(tcb);
1180 
1181 		tcb = (tx_control_block_t *)
1182 		    LIST_GET_NEXT(&pending_list, &tcb->link);
1183 	}
1184 
1185 	/*
1186 	 * Add the tx control blocks in the pending list to the free list.
1187 	 */
1188 	igb_put_free_list(tx_ring, &pending_list);
1189 
1190 	return (desc_num);
1191 }
1192 
1193 /*
1194  * igb_tx_recycle_head_wb
1195  *
1196  * Check the head write-back, and recycle all the transmitted
1197  * tx descriptors and tx control blocks.
1198  */
1199 uint32_t
1200 igb_tx_recycle_head_wb(igb_tx_ring_t *tx_ring)
1201 {
1202 	uint32_t index;
1203 	uint32_t head_wb;
1204 	int desc_num;
1205 	tx_control_block_t *tcb;
1206 	link_list_t pending_list;
1207 	igb_t *igb = tx_ring->igb;
1208 
1209 	/*
1210 	 * The mutex_tryenter() is used to avoid unnecessary
1211 	 * lock contention.
1212 	 */
1213 	if (mutex_tryenter(&tx_ring->recycle_lock) == 0)
1214 		return (0);
1215 
1216 	ASSERT(tx_ring->tbd_free <= tx_ring->ring_size);
1217 
1218 	if (tx_ring->tbd_free == tx_ring->ring_size) {
1219 		tx_ring->recycle_fail = 0;
1220 		tx_ring->stall_watchdog = 0;
1221 		mutex_exit(&tx_ring->recycle_lock);
1222 		return (0);
1223 	}
1224 
1225 	/*
1226 	 * Sync the DMA buffer of the tx descriptor ring
1227 	 *
1228 	 * Note: For head write-back mode, the tx descriptors will not
1229 	 * be written back, but the head write-back value is stored at
1230 	 * the last extra tbd at the end of the DMA area, we still need
1231 	 * to sync the head write-back value for kernel.
1232 	 *
1233 	 * DMA_SYNC(&tx_ring->tbd_area, DDI_DMA_SYNC_FORKERNEL);
1234 	 */
1235 	(void) ddi_dma_sync(tx_ring->tbd_area.dma_handle,
1236 	    sizeof (union e1000_adv_tx_desc) * tx_ring->ring_size,
1237 	    sizeof (uint32_t),
1238 	    DDI_DMA_SYNC_FORKERNEL);
1239 
1240 	if (igb_check_dma_handle(
1241 	    tx_ring->tbd_area.dma_handle) != DDI_FM_OK) {
1242 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
1243 		atomic_or_32(&igb->igb_state, IGB_ERROR);
1244 		return (0);
1245 	}
1246 
1247 	LINK_LIST_INIT(&pending_list);
1248 	desc_num = 0;
1249 	index = tx_ring->tbd_head;	/* Next index to clean */
1250 
1251 	/*
1252 	 * Get the value of head write-back
1253 	 */
1254 	head_wb = *tx_ring->tbd_head_wb;
1255 	while (index != head_wb) {
1256 		tcb = tx_ring->work_list[index];
1257 		ASSERT(tcb != NULL);
1258 
1259 		if (OFFSET(index, head_wb, tx_ring->ring_size) <
1260 		    tcb->desc_num) {
1261 			/*
1262 			 * The current tx control block is not
1263 			 * completely transmitted, stop recycling
1264 			 */
1265 			break;
1266 		}
1267 
1268 		/*
1269 		 * Strip off the tx control block from the work list,
1270 		 * and add it to the pending list.
1271 		 */
1272 		tx_ring->work_list[index] = NULL;
1273 		LIST_PUSH_TAIL(&pending_list, &tcb->link);
1274 
1275 		/*
1276 		 * Advance the index of the tx descriptor ring
1277 		 */
1278 		index = NEXT_INDEX(index, tcb->desc_num, tx_ring->ring_size);
1279 
1280 		/*
1281 		 * Count the total number of the tx descriptors recycled
1282 		 */
1283 		desc_num += tcb->desc_num;
1284 	}
1285 
1286 	/*
1287 	 * If no tx descriptors are recycled, no need to do more processing
1288 	 */
1289 	if (desc_num == 0) {
1290 		tx_ring->recycle_fail++;
1291 		mutex_exit(&tx_ring->recycle_lock);
1292 		return (0);
1293 	}
1294 
1295 	tx_ring->recycle_fail = 0;
1296 	tx_ring->stall_watchdog = 0;
1297 
1298 	/*
1299 	 * Update the head index of the tx descriptor ring
1300 	 */
1301 	tx_ring->tbd_head = index;
1302 
1303 	/*
1304 	 * Update the number of the free tx descriptors with atomic operations
1305 	 */
1306 	atomic_add_32(&tx_ring->tbd_free, desc_num);
1307 
1308 	mutex_exit(&tx_ring->recycle_lock);
1309 
1310 	/*
1311 	 * Free the resources used by the tx control blocks
1312 	 * in the pending list
1313 	 */
1314 	tcb = (tx_control_block_t *)LIST_GET_HEAD(&pending_list);
1315 	while (tcb) {
1316 		/*
1317 		 * Release the resources occupied by the tx control block
1318 		 */
1319 		igb_free_tcb(tcb);
1320 
1321 		tcb = (tx_control_block_t *)
1322 		    LIST_GET_NEXT(&pending_list, &tcb->link);
1323 	}
1324 
1325 	/*
1326 	 * Add the tx control blocks in the pending list to the free list.
1327 	 */
1328 	igb_put_free_list(tx_ring, &pending_list);
1329 
1330 	return (desc_num);
1331 }
1332 
1333 /*
1334  * igb_free_tcb - free up the tx control block
1335  *
1336  * Free the resources of the tx control block, including
1337  * unbind the previously bound DMA handle, and reset other
1338  * control fields.
1339  */
1340 void
1341 igb_free_tcb(tx_control_block_t *tcb)
1342 {
1343 	switch (tcb->tx_type) {
1344 	case USE_COPY:
1345 		/*
1346 		 * Reset the buffer length that is used for copy
1347 		 */
1348 		tcb->tx_buf.len = 0;
1349 		break;
1350 	case USE_DMA:
1351 		/*
1352 		 * Release the DMA resource that is used for
1353 		 * DMA binding.
1354 		 */
1355 		(void) ddi_dma_unbind_handle(tcb->tx_dma_handle);
1356 		break;
1357 	default:
1358 		break;
1359 	}
1360 
1361 	/*
1362 	 * Free the mblk
1363 	 */
1364 	if (tcb->mp != NULL) {
1365 		freemsg(tcb->mp);
1366 		tcb->mp = NULL;
1367 	}
1368 
1369 	tcb->tx_type = USE_NONE;
1370 	tcb->frag_num = 0;
1371 	tcb->desc_num = 0;
1372 }
1373 
1374 /*
1375  * igb_get_free_list - Get a free tx control block from the free list
1376  *
1377  * The atomic operation on the number of the available tx control block
1378  * in the free list is used to keep this routine mutual exclusive with
1379  * the routine igb_put_check_list.
1380  */
1381 static tx_control_block_t *
1382 igb_get_free_list(igb_tx_ring_t *tx_ring)
1383 {
1384 	tx_control_block_t *tcb;
1385 
1386 	/*
1387 	 * Check and update the number of the free tx control block
1388 	 * in the free list.
1389 	 */
1390 	if (igb_atomic_reserve(&tx_ring->tcb_free, 1) < 0)
1391 		return (NULL);
1392 
1393 	mutex_enter(&tx_ring->tcb_head_lock);
1394 
1395 	tcb = tx_ring->free_list[tx_ring->tcb_head];
1396 	ASSERT(tcb != NULL);
1397 	tx_ring->free_list[tx_ring->tcb_head] = NULL;
1398 	tx_ring->tcb_head = NEXT_INDEX(tx_ring->tcb_head, 1,
1399 	    tx_ring->free_list_size);
1400 
1401 	mutex_exit(&tx_ring->tcb_head_lock);
1402 
1403 	return (tcb);
1404 }
1405 
1406 /*
1407  * igb_put_free_list
1408  *
1409  * Put a list of used tx control blocks back to the free list
1410  *
1411  * A mutex is used here to ensure the serialization. The mutual exclusion
1412  * between igb_get_free_list and igb_put_free_list is implemented with
1413  * the atomic operation on the counter tcb_free.
1414  */
1415 void
1416 igb_put_free_list(igb_tx_ring_t *tx_ring, link_list_t *pending_list)
1417 {
1418 	uint32_t index;
1419 	int tcb_num;
1420 	tx_control_block_t *tcb;
1421 
1422 	mutex_enter(&tx_ring->tcb_tail_lock);
1423 
1424 	index = tx_ring->tcb_tail;
1425 
1426 	tcb_num = 0;
1427 	tcb = (tx_control_block_t *)LIST_POP_HEAD(pending_list);
1428 	while (tcb != NULL) {
1429 		ASSERT(tx_ring->free_list[index] == NULL);
1430 		tx_ring->free_list[index] = tcb;
1431 
1432 		tcb_num++;
1433 
1434 		index = NEXT_INDEX(index, 1, tx_ring->free_list_size);
1435 
1436 		tcb = (tx_control_block_t *)LIST_POP_HEAD(pending_list);
1437 	}
1438 
1439 	tx_ring->tcb_tail = index;
1440 
1441 	/*
1442 	 * Update the number of the free tx control block
1443 	 * in the free list. This operation must be placed
1444 	 * under the protection of the lock.
1445 	 */
1446 	atomic_add_32(&tx_ring->tcb_free, tcb_num);
1447 
1448 	mutex_exit(&tx_ring->tcb_tail_lock);
1449 }
1450