xref: /titanic_41/usr/src/uts/common/io/e1000g/e1000g_rx.c (revision 0bb073995ac5a95bd35f2dd790df1ea3d8c2d507)
1 /*
2  * This file is provided under a CDDLv1 license.  When using or
3  * redistributing this file, you may do so under this license.
4  * In redistributing this file this license must be included
5  * and no other modification of this header file is permitted.
6  *
7  * CDDL LICENSE SUMMARY
8  *
9  * Copyright(c) 1999 - 2008 Intel Corporation. All rights reserved.
10  *
11  * The contents of this file are subject to the terms of Version
12  * 1.0 of the Common Development and Distribution License (the "License").
13  *
14  * You should have received a copy of the License with this software.
15  * You can obtain a copy of the License at
16  *	http://www.opensolaris.org/os/licensing.
17  * See the License for the specific language governing permissions
18  * and limitations under the License.
19  */
20 
21 /*
22  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms of the CDDLv1.
24  */
25 
26 /*
27  * **********************************************************************
28  *									*
29  * Module Name:								*
30  *   e1000g_rx.c							*
31  *									*
32  * Abstract:								*
33  *   This file contains some routines that take care of Receive		*
34  *   interrupt and also for the received packets it sends up to		*
35  *   upper layer.							*
36  *   It tries to do a zero copy if free buffers are available in	*
37  *   the pool.								*
38  *									*
39  * **********************************************************************
40  */
41 
42 #include "e1000g_sw.h"
43 #include "e1000g_debug.h"
44 
45 static p_rx_sw_packet_t e1000g_get_buf(e1000g_rx_ring_t *rx_ring);
46 #pragma	inline(e1000g_get_buf)
47 static void e1000g_priv_devi_list_clean();
48 
49 /*
50  * e1000g_rxfree_func - the call-back function to reclaim rx buffer
51  *
52  * This function is called when an mp is freed by the user thru
53  * freeb call (Only for mp constructed through desballoc call)
54  * It returns back the freed buffer to the freelist
55  */
56 void
57 e1000g_rxfree_func(p_rx_sw_packet_t packet)
58 {
59 	e1000g_rx_ring_t *rx_ring;
60 
61 	rx_ring = (e1000g_rx_ring_t *)(uintptr_t)packet->rx_ring;
62 
63 	/*
64 	 * Here the rx recycling processes different rx packets in different
65 	 * threads, so we protect it with RW_READER to ensure it won't block
66 	 * other rx recycling threads.
67 	 */
68 	rw_enter(&e1000g_rx_detach_lock, RW_READER);
69 
70 	if (packet->flag == E1000G_RX_SW_FREE) {
71 		rw_exit(&e1000g_rx_detach_lock);
72 		return;
73 	}
74 
75 	if (packet->flag == E1000G_RX_SW_STOP) {
76 		packet->flag = E1000G_RX_SW_FREE;
77 		rw_exit(&e1000g_rx_detach_lock);
78 
79 		rw_enter(&e1000g_rx_detach_lock, RW_WRITER);
80 		rx_ring->pending_count--;
81 		e1000g_mblks_pending--;
82 
83 		if (rx_ring->pending_count == 0) {
84 			while (rx_ring->pending_list != NULL) {
85 				packet = rx_ring->pending_list;
86 				rx_ring->pending_list =
87 				    rx_ring->pending_list->next;
88 
89 				ASSERT(packet->mp == NULL);
90 				e1000g_free_rx_sw_packet(packet);
91 			}
92 		}
93 
94 		/*
95 		 * If e1000g_force_detach is enabled, we need to clean up
96 		 * the idle priv_dip entries in the private dip list while
97 		 * e1000g_mblks_pending is zero.
98 		 */
99 		if (e1000g_force_detach && (e1000g_mblks_pending == 0))
100 			e1000g_priv_devi_list_clean();
101 		rw_exit(&e1000g_rx_detach_lock);
102 		return;
103 	}
104 
105 	if (packet->flag == E1000G_RX_SW_DETACH) {
106 		packet->flag = E1000G_RX_SW_FREE;
107 		rw_exit(&e1000g_rx_detach_lock);
108 
109 		ASSERT(packet->mp == NULL);
110 		e1000g_free_rx_sw_packet(packet);
111 
112 		/*
113 		 * Here the e1000g_mblks_pending may be modified by different
114 		 * rx recycling threads simultaneously, so we need to protect
115 		 * it with RW_WRITER.
116 		 */
117 		rw_enter(&e1000g_rx_detach_lock, RW_WRITER);
118 		e1000g_mblks_pending--;
119 
120 		/*
121 		 * If e1000g_force_detach is enabled, we need to clean up
122 		 * the idle priv_dip entries in the private dip list while
123 		 * e1000g_mblks_pending is zero.
124 		 */
125 		if (e1000g_force_detach && (e1000g_mblks_pending == 0))
126 			e1000g_priv_devi_list_clean();
127 		rw_exit(&e1000g_rx_detach_lock);
128 		return;
129 	}
130 
131 	packet->flag = E1000G_RX_SW_FREE;
132 
133 	if (packet->mp == NULL) {
134 		/*
135 		 * Allocate a mblk that binds to the data buffer
136 		 */
137 		packet->mp = desballoc((unsigned char *)
138 		    packet->rx_buf->address - E1000G_IPALIGNROOM,
139 		    packet->rx_buf->size + E1000G_IPALIGNROOM,
140 		    BPRI_MED, &packet->free_rtn);
141 
142 		if (packet->mp != NULL) {
143 			packet->mp->b_rptr += E1000G_IPALIGNROOM;
144 			packet->mp->b_wptr += E1000G_IPALIGNROOM;
145 		} else {
146 			E1000G_STAT(rx_ring->stat_esballoc_fail);
147 		}
148 	}
149 
150 	mutex_enter(&rx_ring->freelist_lock);
151 	QUEUE_PUSH_TAIL(&rx_ring->free_list, &packet->Link);
152 	rx_ring->avail_freepkt++;
153 	mutex_exit(&rx_ring->freelist_lock);
154 
155 	rw_exit(&e1000g_rx_detach_lock);
156 }
157 
158 /*
159  * e1000g_priv_devi_list_clean - clean up e1000g_private_devi_list
160  *
161  * We will walk the e1000g_private_devi_list to free the entry marked
162  * with the E1000G_PRIV_DEVI_DETACH flag.
163  */
164 static void
165 e1000g_priv_devi_list_clean()
166 {
167 	private_devi_list_t *devi_node, *devi_del;
168 
169 	if (e1000g_private_devi_list == NULL)
170 		return;
171 
172 	devi_node = e1000g_private_devi_list;
173 	while ((devi_node != NULL) &&
174 	    (devi_node->flag == E1000G_PRIV_DEVI_DETACH)) {
175 		e1000g_private_devi_list = devi_node->next;
176 		kmem_free(devi_node->priv_dip,
177 		    sizeof (struct dev_info));
178 		kmem_free(devi_node,
179 		    sizeof (private_devi_list_t));
180 		devi_node = e1000g_private_devi_list;
181 	}
182 	if (e1000g_private_devi_list == NULL)
183 		return;
184 	while (devi_node->next != NULL) {
185 		if (devi_node->next->flag == E1000G_PRIV_DEVI_DETACH) {
186 			devi_del = devi_node->next;
187 			devi_node->next = devi_del->next;
188 			kmem_free(devi_del->priv_dip,
189 			    sizeof (struct dev_info));
190 			kmem_free(devi_del,
191 			    sizeof (private_devi_list_t));
192 		} else {
193 			devi_node = devi_node->next;
194 		}
195 	}
196 }
197 
198 /*
199  * e1000g_rx_setup - setup rx data structures
200  *
201  * This routine initializes all of the receive related
202  * structures. This includes the receive descriptors, the
203  * actual receive buffers, and the rx_sw_packet software
204  * structures.
205  */
206 void
207 e1000g_rx_setup(struct e1000g *Adapter)
208 {
209 	struct e1000_hw *hw;
210 	p_rx_sw_packet_t packet;
211 	struct e1000_rx_desc *descriptor;
212 	uint32_t buf_low;
213 	uint32_t buf_high;
214 	uint32_t reg_val;
215 	uint32_t rctl;
216 	uint32_t rxdctl;
217 	uint32_t ert;
218 	int i;
219 	int size;
220 	e1000g_rx_ring_t *rx_ring;
221 
222 	hw = &Adapter->shared;
223 	rx_ring = Adapter->rx_ring;
224 
225 	/*
226 	 * zero out all of the receive buffer descriptor memory
227 	 * assures any previous data or status is erased
228 	 */
229 	bzero(rx_ring->rbd_area,
230 	    sizeof (struct e1000_rx_desc) * Adapter->rx_desc_num);
231 
232 	if (!Adapter->rx_buffer_setup) {
233 		/* Init the list of "Receive Buffer" */
234 		QUEUE_INIT_LIST(&rx_ring->recv_list);
235 
236 		/* Init the list of "Free Receive Buffer" */
237 		QUEUE_INIT_LIST(&rx_ring->free_list);
238 
239 		/*
240 		 * Setup Receive list and the Free list. Note that
241 		 * the both were allocated in one packet area.
242 		 */
243 		packet = rx_ring->packet_area;
244 		descriptor = rx_ring->rbd_first;
245 
246 		for (i = 0; i < Adapter->rx_desc_num;
247 		    i++, packet = packet->next, descriptor++) {
248 			ASSERT(packet != NULL);
249 			ASSERT(descriptor != NULL);
250 			descriptor->buffer_addr =
251 			    packet->rx_buf->dma_address;
252 
253 			/* Add this rx_sw_packet to the receive list */
254 			QUEUE_PUSH_TAIL(&rx_ring->recv_list,
255 			    &packet->Link);
256 		}
257 
258 		for (i = 0; i < Adapter->rx_freelist_num;
259 		    i++, packet = packet->next) {
260 			ASSERT(packet != NULL);
261 			/* Add this rx_sw_packet to the free list */
262 			QUEUE_PUSH_TAIL(&rx_ring->free_list,
263 			    &packet->Link);
264 		}
265 		rx_ring->avail_freepkt = Adapter->rx_freelist_num;
266 
267 		Adapter->rx_buffer_setup = B_TRUE;
268 	} else {
269 		/* Setup the initial pointer to the first rx descriptor */
270 		packet = (p_rx_sw_packet_t)
271 		    QUEUE_GET_HEAD(&rx_ring->recv_list);
272 		descriptor = rx_ring->rbd_first;
273 
274 		for (i = 0; i < Adapter->rx_desc_num; i++) {
275 			ASSERT(packet != NULL);
276 			ASSERT(descriptor != NULL);
277 			descriptor->buffer_addr =
278 			    packet->rx_buf->dma_address;
279 
280 			/* Get next rx_sw_packet */
281 			packet = (p_rx_sw_packet_t)
282 			    QUEUE_GET_NEXT(&rx_ring->recv_list, &packet->Link);
283 			descriptor++;
284 		}
285 	}
286 
287 	E1000_WRITE_REG(&Adapter->shared, E1000_RDTR, Adapter->rx_intr_delay);
288 	E1000G_DEBUGLOG_1(Adapter, E1000G_INFO_LEVEL,
289 	    "E1000_RDTR: 0x%x\n", Adapter->rx_intr_delay);
290 	if (hw->mac.type >= e1000_82540) {
291 		E1000_WRITE_REG(&Adapter->shared, E1000_RADV,
292 		    Adapter->rx_intr_abs_delay);
293 		E1000G_DEBUGLOG_1(Adapter, E1000G_INFO_LEVEL,
294 		    "E1000_RADV: 0x%x\n", Adapter->rx_intr_abs_delay);
295 	}
296 
297 	/*
298 	 * Setup our descriptor pointers
299 	 */
300 	rx_ring->rbd_next = rx_ring->rbd_first;
301 
302 	size = Adapter->rx_desc_num * sizeof (struct e1000_rx_desc);
303 	E1000_WRITE_REG(hw, E1000_RDLEN(0), size);
304 	size = E1000_READ_REG(hw, E1000_RDLEN(0));
305 
306 	/* To get lower order bits */
307 	buf_low = (uint32_t)rx_ring->rbd_dma_addr;
308 	/* To get the higher order bits */
309 	buf_high = (uint32_t)(rx_ring->rbd_dma_addr >> 32);
310 
311 	E1000_WRITE_REG(hw, E1000_RDBAH(0), buf_high);
312 	E1000_WRITE_REG(hw, E1000_RDBAL(0), buf_low);
313 
314 	/*
315 	 * Setup our HW Rx Head & Tail descriptor pointers
316 	 */
317 	E1000_WRITE_REG(hw, E1000_RDT(0),
318 	    (uint32_t)(rx_ring->rbd_last - rx_ring->rbd_first));
319 	E1000_WRITE_REG(hw, E1000_RDH(0), 0);
320 
321 	/*
322 	 * Setup the Receive Control Register (RCTL), and ENABLE the
323 	 * receiver. The initial configuration is to: Enable the receiver,
324 	 * accept broadcasts, discard bad packets (and long packets),
325 	 * disable VLAN filter checking, set the receive descriptor
326 	 * minimum threshold size to 1/2, and the receive buffer size to
327 	 * 2k.
328 	 */
329 	rctl = E1000_RCTL_EN |		/* Enable Receive Unit */
330 	    E1000_RCTL_BAM |		/* Accept Broadcast Packets */
331 	    E1000_RCTL_LPE |		/* Large Packet Enable bit */
332 	    (hw->mac.mc_filter_type << E1000_RCTL_MO_SHIFT) |
333 	    E1000_RCTL_RDMTS_HALF |
334 	    E1000_RCTL_LBM_NO;		/* Loopback Mode = none */
335 
336 	if (Adapter->strip_crc)
337 		rctl |= E1000_RCTL_SECRC;	/* Strip Ethernet CRC */
338 
339 	if ((Adapter->max_frame_size > FRAME_SIZE_UPTO_2K) &&
340 	    (Adapter->max_frame_size <= FRAME_SIZE_UPTO_4K))
341 		rctl |= E1000_RCTL_SZ_4096 | E1000_RCTL_BSEX;
342 	else if ((Adapter->max_frame_size > FRAME_SIZE_UPTO_4K) &&
343 	    (Adapter->max_frame_size <= FRAME_SIZE_UPTO_8K))
344 		rctl |= E1000_RCTL_SZ_8192 | E1000_RCTL_BSEX;
345 	else if ((Adapter->max_frame_size > FRAME_SIZE_UPTO_8K) &&
346 	    (Adapter->max_frame_size <= FRAME_SIZE_UPTO_16K))
347 		rctl |= E1000_RCTL_SZ_16384 | E1000_RCTL_BSEX;
348 	else
349 		rctl |= E1000_RCTL_SZ_2048;
350 
351 	if (e1000_tbi_sbp_enabled_82543(hw))
352 		rctl |= E1000_RCTL_SBP;
353 
354 	/*
355 	 * Enable early receives on supported devices, only takes effect when
356 	 * packet size is equal or larger than the specified value (in 8 byte
357 	 * units), e.g. using jumbo frames when setting to E1000_ERT_2048
358 	 */
359 	if ((hw->mac.type == e1000_82573) ||
360 	    (hw->mac.type == e1000_82574) ||
361 	    (hw->mac.type == e1000_ich9lan) ||
362 	    (hw->mac.type == e1000_ich10lan)) {
363 
364 		ert = E1000_ERT_2048;
365 
366 		/*
367 		 * Special modification when ERT and
368 		 * jumbo frames are enabled
369 		 */
370 		if (Adapter->default_mtu > ETHERMTU) {
371 			rxdctl = E1000_READ_REG(hw, E1000_RXDCTL(0));
372 			E1000_WRITE_REG(hw, E1000_RXDCTL(0), rxdctl | 0x3);
373 			ert |= (1 << 13);
374 		}
375 
376 		E1000_WRITE_REG(hw, E1000_ERT, ert);
377 	}
378 
379 	reg_val =
380 	    E1000_RXCSUM_TUOFL |	/* TCP/UDP checksum offload Enable */
381 	    E1000_RXCSUM_IPOFL;		/* IP checksum offload Enable */
382 
383 	E1000_WRITE_REG(hw, E1000_RXCSUM, reg_val);
384 
385 	/*
386 	 * Workaround: Set bit 16 (IPv6_ExDIS) to disable the
387 	 * processing of received IPV6 extension headers
388 	 */
389 	if ((hw->mac.type == e1000_82571) || (hw->mac.type == e1000_82572)) {
390 		reg_val = E1000_READ_REG(hw, E1000_RFCTL);
391 		reg_val |= (E1000_RFCTL_IPV6_EX_DIS |
392 		    E1000_RFCTL_NEW_IPV6_EXT_DIS);
393 		E1000_WRITE_REG(hw, E1000_RFCTL, reg_val);
394 	}
395 
396 	/* Write to enable the receive unit */
397 	E1000_WRITE_REG(hw, E1000_RCTL, rctl);
398 }
399 
400 /*
401  * e1000g_get_buf - get an rx sw packet from the free_list
402  */
403 static p_rx_sw_packet_t
404 e1000g_get_buf(e1000g_rx_ring_t *rx_ring)
405 {
406 	p_rx_sw_packet_t packet;
407 
408 	mutex_enter(&rx_ring->freelist_lock);
409 	packet = (p_rx_sw_packet_t)
410 	    QUEUE_POP_HEAD(&rx_ring->free_list);
411 	if (packet != NULL)
412 		rx_ring->avail_freepkt--;
413 	mutex_exit(&rx_ring->freelist_lock);
414 
415 	return (packet);
416 }
417 
418 /*
419  * e1000g_receive - main receive routine
420  *
421  * This routine will process packets received in an interrupt
422  */
423 mblk_t *
424 e1000g_receive(struct e1000g *Adapter)
425 {
426 	struct e1000_hw *hw;
427 	mblk_t *nmp;
428 	mblk_t *ret_mp;
429 	mblk_t *ret_nmp;
430 	struct e1000_rx_desc *current_desc;
431 	struct e1000_rx_desc *last_desc;
432 	p_rx_sw_packet_t packet;
433 	p_rx_sw_packet_t newpkt;
434 	uint16_t length;
435 	uint32_t pkt_count;
436 	uint32_t desc_count;
437 	boolean_t accept_frame;
438 	boolean_t end_of_packet;
439 	boolean_t need_copy;
440 	e1000g_rx_ring_t *rx_ring;
441 	dma_buffer_t *rx_buf;
442 	uint16_t cksumflags;
443 
444 	ret_mp = NULL;
445 	ret_nmp = NULL;
446 	pkt_count = 0;
447 	desc_count = 0;
448 	cksumflags = 0;
449 
450 	hw = &Adapter->shared;
451 	rx_ring = Adapter->rx_ring;
452 
453 	/* Sync the Rx descriptor DMA buffers */
454 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
455 	    0, 0, DDI_DMA_SYNC_FORKERNEL);
456 
457 	if (e1000g_check_dma_handle(rx_ring->rbd_dma_handle) != DDI_FM_OK) {
458 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
459 		Adapter->chip_state = E1000G_ERROR;
460 	}
461 
462 	current_desc = rx_ring->rbd_next;
463 	if (!(current_desc->status & E1000_RXD_STAT_DD)) {
464 		/*
465 		 * don't send anything up. just clear the RFD
466 		 */
467 		E1000G_DEBUG_STAT(rx_ring->stat_none);
468 		return (ret_mp);
469 	}
470 
471 	/*
472 	 * Loop through the receive descriptors starting at the last known
473 	 * descriptor owned by the hardware that begins a packet.
474 	 */
475 	while ((current_desc->status & E1000_RXD_STAT_DD) &&
476 	    (pkt_count < Adapter->rx_limit_onintr)) {
477 
478 		desc_count++;
479 		/*
480 		 * Now this can happen in Jumbo frame situation.
481 		 */
482 		if (current_desc->status & E1000_RXD_STAT_EOP) {
483 			/* packet has EOP set */
484 			end_of_packet = B_TRUE;
485 		} else {
486 			/*
487 			 * If this received buffer does not have the
488 			 * End-Of-Packet bit set, the received packet
489 			 * will consume multiple buffers. We won't send this
490 			 * packet upstack till we get all the related buffers.
491 			 */
492 			end_of_packet = B_FALSE;
493 		}
494 
495 		/*
496 		 * Get a pointer to the actual receive buffer
497 		 * The mp->b_rptr is mapped to The CurrentDescriptor
498 		 * Buffer Address.
499 		 */
500 		packet =
501 		    (p_rx_sw_packet_t)QUEUE_GET_HEAD(&rx_ring->recv_list);
502 		ASSERT(packet != NULL);
503 
504 		rx_buf = packet->rx_buf;
505 
506 		length = current_desc->length;
507 
508 #ifdef __sparc
509 		if (packet->dma_type == USE_DVMA)
510 			dvma_sync(rx_buf->dma_handle, 0,
511 			    DDI_DMA_SYNC_FORKERNEL);
512 		else
513 			(void) ddi_dma_sync(rx_buf->dma_handle,
514 			    E1000G_IPALIGNROOM, length,
515 			    DDI_DMA_SYNC_FORKERNEL);
516 #else
517 		(void) ddi_dma_sync(rx_buf->dma_handle,
518 		    E1000G_IPALIGNROOM, length,
519 		    DDI_DMA_SYNC_FORKERNEL);
520 #endif
521 
522 		if (e1000g_check_dma_handle(
523 		    rx_buf->dma_handle) != DDI_FM_OK) {
524 			ddi_fm_service_impact(Adapter->dip,
525 			    DDI_SERVICE_DEGRADED);
526 			Adapter->chip_state = E1000G_ERROR;
527 		}
528 
529 		accept_frame = (current_desc->errors == 0) ||
530 		    ((current_desc->errors &
531 		    (E1000_RXD_ERR_TCPE | E1000_RXD_ERR_IPE)) != 0);
532 
533 		if (hw->mac.type == e1000_82543) {
534 			unsigned char last_byte;
535 
536 			last_byte =
537 			    *((unsigned char *)rx_buf->address + length - 1);
538 
539 			if (TBI_ACCEPT(hw,
540 			    current_desc->status, current_desc->errors,
541 			    current_desc->length, last_byte,
542 			    Adapter->min_frame_size, Adapter->max_frame_size)) {
543 
544 				e1000_tbi_adjust_stats(Adapter,
545 				    length, hw->mac.addr);
546 
547 				length--;
548 				accept_frame = B_TRUE;
549 			} else if (e1000_tbi_sbp_enabled_82543(hw) &&
550 			    (current_desc->errors == E1000_RXD_ERR_CE)) {
551 				accept_frame = B_TRUE;
552 			}
553 		}
554 
555 		/*
556 		 * Indicate the packet to the NOS if it was good.
557 		 * Normally, hardware will discard bad packets for us.
558 		 * Check for the packet to be a valid Ethernet packet
559 		 */
560 		if (!accept_frame) {
561 			/*
562 			 * error in incoming packet, either the packet is not a
563 			 * ethernet size packet, or the packet has an error. In
564 			 * either case, the packet will simply be discarded.
565 			 */
566 			E1000G_DEBUGLOG_0(Adapter, E1000G_INFO_LEVEL,
567 			    "Process Receive Interrupts: Error in Packet\n");
568 
569 			E1000G_STAT(rx_ring->stat_error);
570 			/*
571 			 * Returning here as we are done here. There is
572 			 * no point in waiting for while loop to elapse
573 			 * and the things which were done. More efficient
574 			 * and less error prone...
575 			 */
576 			goto rx_drop;
577 		}
578 
579 		/*
580 		 * If the Ethernet CRC is not stripped by the hardware,
581 		 * we need to strip it before sending it up to the stack.
582 		 */
583 		if (end_of_packet && !Adapter->strip_crc) {
584 			if (length > ETHERFCSL) {
585 				length -= ETHERFCSL;
586 			} else {
587 				/*
588 				 * If the fragment is smaller than the CRC,
589 				 * drop this fragment, do the processing of
590 				 * the end of the packet.
591 				 */
592 				ASSERT(rx_ring->rx_mblk_tail != NULL);
593 				rx_ring->rx_mblk_tail->b_wptr -=
594 				    ETHERFCSL - length;
595 				rx_ring->rx_mblk_len -=
596 				    ETHERFCSL - length;
597 
598 				QUEUE_POP_HEAD(&rx_ring->recv_list);
599 
600 				goto rx_end_of_packet;
601 			}
602 		}
603 
604 		need_copy = B_TRUE;
605 
606 		if (length <= Adapter->rx_bcopy_thresh)
607 			goto rx_copy;
608 
609 		/*
610 		 * Get the pre-constructed mblk that was associated
611 		 * to the receive data buffer.
612 		 */
613 		if (packet->mp == NULL) {
614 			packet->mp = desballoc((unsigned char *)
615 			    rx_buf->address - E1000G_IPALIGNROOM,
616 			    length + E1000G_IPALIGNROOM,
617 			    BPRI_MED, &packet->free_rtn);
618 
619 			if (packet->mp != NULL) {
620 				packet->mp->b_rptr += E1000G_IPALIGNROOM;
621 				packet->mp->b_wptr += E1000G_IPALIGNROOM;
622 			} else {
623 				E1000G_STAT(rx_ring->stat_esballoc_fail);
624 			}
625 		}
626 
627 		if (packet->mp != NULL) {
628 			/*
629 			 * We have two sets of buffer pool. One associated with
630 			 * the Rxdescriptors and other a freelist buffer pool.
631 			 * Each time we get a good packet, Try to get a buffer
632 			 * from the freelist pool using e1000g_get_buf. If we
633 			 * get free buffer, then replace the descriptor buffer
634 			 * address with the free buffer we just got, and pass
635 			 * the pre-constructed mblk upstack. (note no copying)
636 			 *
637 			 * If we failed to get a free buffer, then try to
638 			 * allocate a new buffer(mp) and copy the recv buffer
639 			 * content to our newly allocated buffer(mp). Don't
640 			 * disturb the desriptor buffer address. (note copying)
641 			 */
642 			newpkt = e1000g_get_buf(rx_ring);
643 
644 			if (newpkt != NULL) {
645 				/*
646 				 * Get the mblk associated to the data,
647 				 * and strip it off the sw packet.
648 				 */
649 				nmp = packet->mp;
650 				packet->mp = NULL;
651 				packet->flag = E1000G_RX_SW_SENDUP;
652 
653 				/*
654 				 * Now replace old buffer with the new
655 				 * one we got from free list
656 				 * Both the RxSwPacket as well as the
657 				 * Receive Buffer Descriptor will now
658 				 * point to this new packet.
659 				 */
660 				packet = newpkt;
661 
662 				current_desc->buffer_addr =
663 				    newpkt->rx_buf->dma_address;
664 
665 				need_copy = B_FALSE;
666 			} else {
667 				E1000G_DEBUG_STAT(rx_ring->stat_no_freepkt);
668 			}
669 		}
670 
671 rx_copy:
672 		if (need_copy) {
673 			/*
674 			 * No buffers available on free list,
675 			 * bcopy the data from the buffer and
676 			 * keep the original buffer. Dont want to
677 			 * do this.. Yack but no other way
678 			 */
679 			if ((nmp = allocb(length + E1000G_IPALIGNROOM,
680 			    BPRI_MED)) == NULL) {
681 				/*
682 				 * The system has no buffers available
683 				 * to send up the incoming packet, hence
684 				 * the packet will have to be processed
685 				 * when there're more buffers available.
686 				 */
687 				E1000G_STAT(rx_ring->stat_allocb_fail);
688 				goto rx_drop;
689 			}
690 			nmp->b_rptr += E1000G_IPALIGNROOM;
691 			nmp->b_wptr += E1000G_IPALIGNROOM;
692 			/*
693 			 * The free list did not have any buffers
694 			 * available, so, the received packet will
695 			 * have to be copied into a mp and the original
696 			 * buffer will have to be retained for future
697 			 * packet reception.
698 			 */
699 			bcopy(rx_buf->address, nmp->b_wptr, length);
700 		}
701 
702 		/*
703 		 * The rx_sw_packet MUST be popped off the
704 		 * RxSwPacketList before either a putnext or freemsg
705 		 * is done on the mp that has now been created by the
706 		 * desballoc. If not, it is possible that the free
707 		 * routine will get called from the interrupt context
708 		 * and try to put this packet on the free list
709 		 */
710 		(p_rx_sw_packet_t)QUEUE_POP_HEAD(&rx_ring->recv_list);
711 
712 		ASSERT(nmp != NULL);
713 		nmp->b_wptr += length;
714 
715 		if (rx_ring->rx_mblk == NULL) {
716 			/*
717 			 *  TCP/UDP checksum offload and
718 			 *  IP checksum offload
719 			 */
720 			if (!(current_desc->status & E1000_RXD_STAT_IXSM)) {
721 				/*
722 				 * Check TCP/UDP checksum
723 				 */
724 				if ((current_desc->status &
725 				    E1000_RXD_STAT_TCPCS) &&
726 				    !(current_desc->errors &
727 				    E1000_RXD_ERR_TCPE))
728 					cksumflags |= HCK_FULLCKSUM |
729 					    HCK_FULLCKSUM_OK;
730 				/*
731 				 * Check IP Checksum
732 				 */
733 				if ((current_desc->status &
734 				    E1000_RXD_STAT_IPCS) &&
735 				    !(current_desc->errors &
736 				    E1000_RXD_ERR_IPE))
737 					cksumflags |= HCK_IPV4_HDRCKSUM;
738 			}
739 		}
740 
741 		/*
742 		 * We need to maintain our packet chain in the global
743 		 * Adapter structure, for the Rx processing can end
744 		 * with a fragment that has no EOP set.
745 		 */
746 		if (rx_ring->rx_mblk == NULL) {
747 			/* Get the head of the message chain */
748 			rx_ring->rx_mblk = nmp;
749 			rx_ring->rx_mblk_tail = nmp;
750 			rx_ring->rx_mblk_len = length;
751 		} else {	/* Not the first packet */
752 			/* Continue adding buffers */
753 			rx_ring->rx_mblk_tail->b_cont = nmp;
754 			rx_ring->rx_mblk_tail = nmp;
755 			rx_ring->rx_mblk_len += length;
756 		}
757 		ASSERT(rx_ring->rx_mblk != NULL);
758 		ASSERT(rx_ring->rx_mblk_tail != NULL);
759 		ASSERT(rx_ring->rx_mblk_tail->b_cont == NULL);
760 
761 		/*
762 		 * Now this MP is ready to travel upwards but some more
763 		 * fragments are coming.
764 		 * We will send packet upwards as soon as we get EOP
765 		 * set on the packet.
766 		 */
767 		if (!end_of_packet) {
768 			/*
769 			 * continue to get the next descriptor,
770 			 * Tail would be advanced at the end
771 			 */
772 			goto rx_next_desc;
773 		}
774 
775 rx_end_of_packet:
776 		/*
777 		 * Found packet with EOP
778 		 * Process the last fragment.
779 		 */
780 		if (cksumflags != 0) {
781 			(void) hcksum_assoc(rx_ring->rx_mblk,
782 			    NULL, NULL, 0, 0, 0, 0, cksumflags, 0);
783 			cksumflags = 0;
784 		}
785 
786 		/*
787 		 * Count packets that span multi-descriptors
788 		 */
789 		E1000G_DEBUG_STAT_COND(rx_ring->stat_multi_desc,
790 		    (rx_ring->rx_mblk->b_cont != NULL));
791 
792 		/*
793 		 * Append to list to send upstream
794 		 */
795 		if (ret_mp == NULL) {
796 			ret_mp = ret_nmp = rx_ring->rx_mblk;
797 		} else {
798 			ret_nmp->b_next = rx_ring->rx_mblk;
799 			ret_nmp = rx_ring->rx_mblk;
800 		}
801 		ret_nmp->b_next = NULL;
802 
803 		rx_ring->rx_mblk = NULL;
804 		rx_ring->rx_mblk_tail = NULL;
805 		rx_ring->rx_mblk_len = 0;
806 
807 		pkt_count++;
808 
809 rx_next_desc:
810 		/*
811 		 * Zero out the receive descriptors status
812 		 */
813 		current_desc->status = 0;
814 
815 		if (current_desc == rx_ring->rbd_last)
816 			rx_ring->rbd_next = rx_ring->rbd_first;
817 		else
818 			rx_ring->rbd_next++;
819 
820 		last_desc = current_desc;
821 		current_desc = rx_ring->rbd_next;
822 
823 		/*
824 		 * Put the buffer that we just indicated back
825 		 * at the end of our list
826 		 */
827 		QUEUE_PUSH_TAIL(&rx_ring->recv_list,
828 		    &packet->Link);
829 	}	/* while loop */
830 
831 	/* Sync the Rx descriptor DMA buffers */
832 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
833 	    0, 0, DDI_DMA_SYNC_FORDEV);
834 
835 	/*
836 	 * Advance the E1000's Receive Queue #0 "Tail Pointer".
837 	 */
838 	E1000_WRITE_REG(hw, E1000_RDT(0),
839 	    (uint32_t)(last_desc - rx_ring->rbd_first));
840 
841 	if (e1000g_check_acc_handle(Adapter->osdep.reg_handle) != DDI_FM_OK) {
842 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
843 		Adapter->chip_state = E1000G_ERROR;
844 	}
845 
846 	Adapter->rx_pkt_cnt = pkt_count;
847 
848 	return (ret_mp);
849 
850 rx_drop:
851 	/*
852 	 * Zero out the receive descriptors status
853 	 */
854 	current_desc->status = 0;
855 
856 	/* Sync the Rx descriptor DMA buffers */
857 	(void) ddi_dma_sync(rx_ring->rbd_dma_handle,
858 	    0, 0, DDI_DMA_SYNC_FORDEV);
859 
860 	if (current_desc == rx_ring->rbd_last)
861 		rx_ring->rbd_next = rx_ring->rbd_first;
862 	else
863 		rx_ring->rbd_next++;
864 
865 	last_desc = current_desc;
866 
867 	(p_rx_sw_packet_t)QUEUE_POP_HEAD(&rx_ring->recv_list);
868 
869 	QUEUE_PUSH_TAIL(&rx_ring->recv_list, &packet->Link);
870 	/*
871 	 * Reclaim all old buffers already allocated during
872 	 * Jumbo receives.....for incomplete reception
873 	 */
874 	if (rx_ring->rx_mblk != NULL) {
875 		freemsg(rx_ring->rx_mblk);
876 		rx_ring->rx_mblk = NULL;
877 		rx_ring->rx_mblk_tail = NULL;
878 		rx_ring->rx_mblk_len = 0;
879 	}
880 	/*
881 	 * Advance the E1000's Receive Queue #0 "Tail Pointer".
882 	 */
883 	E1000_WRITE_REG(hw, E1000_RDT(0),
884 	    (uint32_t)(last_desc - rx_ring->rbd_first));
885 
886 	if (e1000g_check_acc_handle(Adapter->osdep.reg_handle) != DDI_FM_OK) {
887 		ddi_fm_service_impact(Adapter->dip, DDI_SERVICE_DEGRADED);
888 		Adapter->chip_state = E1000G_ERROR;
889 	}
890 
891 	return (ret_mp);
892 }
893