xref: /titanic_41/usr/src/uts/common/io/igb/igb_rx.c (revision 26fd77009b17f8c8fb32eb362584cfd635e87ad9)
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 2009 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms of the CDDL.
27  */
28 
29 #include "igb_sw.h"
30 
31 /* function prototypes */
32 static mblk_t *igb_rx_bind(igb_rx_ring_t *, uint32_t, uint32_t);
33 static mblk_t *igb_rx_copy(igb_rx_ring_t *, uint32_t, uint32_t);
34 static void igb_rx_assoc_hcksum(mblk_t *, uint32_t);
35 
36 #ifndef IGB_DEBUG
37 #pragma inline(igb_rx_assoc_hcksum)
38 #endif
39 
40 
41 /*
42  * igb_rx_recycle - the call-back function to reclaim rx buffer
43  *
44  * This function is called when an mp is freed by the user thru
45  * freeb call (Only for mp constructed through desballoc call).
46  * It returns back the freed buffer to the free list.
47  */
48 void
49 igb_rx_recycle(caddr_t arg)
50 {
51 	igb_rx_ring_t *rx_ring;
52 	rx_control_block_t *recycle_rcb;
53 	uint32_t free_index;
54 
55 	recycle_rcb = (rx_control_block_t *)(uintptr_t)arg;
56 	rx_ring = recycle_rcb->rx_ring;
57 
58 	if (recycle_rcb->state == RCB_FREE)
59 		return;
60 
61 	recycle_rcb->state = RCB_FREE;
62 
63 	ASSERT(recycle_rcb->mp == NULL);
64 
65 	/*
66 	 * Using the recycled data buffer to generate a new mblk
67 	 */
68 	recycle_rcb->mp = desballoc((unsigned char *)
69 	    recycle_rcb->rx_buf.address,
70 	    recycle_rcb->rx_buf.size,
71 	    0, &recycle_rcb->free_rtn);
72 
73 	/*
74 	 * Put the recycled rx control block into free list
75 	 */
76 	mutex_enter(&rx_ring->recycle_lock);
77 
78 	free_index = rx_ring->rcb_tail;
79 	ASSERT(rx_ring->free_list[free_index] == NULL);
80 
81 	rx_ring->free_list[free_index] = recycle_rcb;
82 	rx_ring->rcb_tail = NEXT_INDEX(free_index, 1, rx_ring->free_list_size);
83 
84 	mutex_exit(&rx_ring->recycle_lock);
85 
86 	/*
87 	 * The atomic operation on the number of the available rx control
88 	 * blocks in the free list is used to make the recycling mutual
89 	 * exclusive with the receiving.
90 	 */
91 	atomic_inc_32(&rx_ring->rcb_free);
92 	ASSERT(rx_ring->rcb_free <= rx_ring->free_list_size);
93 }
94 
95 /*
96  * igb_rx_copy - Use copy to process the received packet
97  *
98  * This function will use bcopy to process the packet
99  * and send the copied packet upstream
100  */
101 static mblk_t *
102 igb_rx_copy(igb_rx_ring_t *rx_ring, uint32_t index, uint32_t pkt_len)
103 {
104 	rx_control_block_t *current_rcb;
105 	mblk_t *mp;
106 	igb_t *igb = rx_ring->igb;
107 
108 	current_rcb = rx_ring->work_list[index];
109 
110 	DMA_SYNC(&current_rcb->rx_buf, DDI_DMA_SYNC_FORKERNEL);
111 
112 	if (igb_check_dma_handle(
113 	    current_rcb->rx_buf.dma_handle) != DDI_FM_OK) {
114 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
115 	}
116 
117 	/*
118 	 * Allocate buffer to receive this packet
119 	 */
120 	mp = allocb(pkt_len + IPHDR_ALIGN_ROOM, 0);
121 	if (mp == NULL) {
122 		igb_log(rx_ring->igb, "igb_rx_copy: allocate buffer failed");
123 		return (NULL);
124 	}
125 
126 	/*
127 	 * Copy the data received into the new cluster
128 	 */
129 	mp->b_rptr += IPHDR_ALIGN_ROOM;
130 	bcopy(current_rcb->rx_buf.address, mp->b_rptr, pkt_len);
131 	mp->b_wptr = mp->b_rptr + pkt_len;
132 
133 	return (mp);
134 }
135 
136 /*
137  * igb_rx_bind - Use existing DMA buffer to build mblk for receiving
138  *
139  * This function will use pre-bound DMA buffer to receive the packet
140  * and build mblk that will be sent upstream.
141  */
142 static mblk_t *
143 igb_rx_bind(igb_rx_ring_t *rx_ring, uint32_t index, uint32_t pkt_len)
144 {
145 	rx_control_block_t *current_rcb;
146 	rx_control_block_t *free_rcb;
147 	uint32_t free_index;
148 	mblk_t *mp;
149 	igb_t *igb = rx_ring->igb;
150 
151 	/*
152 	 * If the free list is empty, we cannot proceed to send
153 	 * the current DMA buffer upstream. We'll have to return
154 	 * and use bcopy to process the packet.
155 	 */
156 	if (igb_atomic_reserve(&rx_ring->rcb_free, 1) < 0)
157 		return (NULL);
158 
159 	current_rcb = rx_ring->work_list[index];
160 	/*
161 	 * If the mp of the rx control block is NULL, try to do
162 	 * desballoc again.
163 	 */
164 	if (current_rcb->mp == NULL) {
165 		current_rcb->mp = desballoc((unsigned char *)
166 		    current_rcb->rx_buf.address,
167 		    current_rcb->rx_buf.size,
168 		    0, &current_rcb->free_rtn);
169 		/*
170 		 * If it is failed to built a mblk using the current
171 		 * DMA buffer, we have to return and use bcopy to
172 		 * process the packet.
173 		 */
174 		if (current_rcb->mp == NULL) {
175 			atomic_inc_32(&rx_ring->rcb_free);
176 			return (NULL);
177 		}
178 	}
179 	/*
180 	 * Sync up the data received
181 	 */
182 	DMA_SYNC(&current_rcb->rx_buf, DDI_DMA_SYNC_FORKERNEL);
183 
184 	if (igb_check_dma_handle(
185 	    current_rcb->rx_buf.dma_handle) != DDI_FM_OK) {
186 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
187 	}
188 
189 	mp = current_rcb->mp;
190 	current_rcb->mp = NULL;
191 	current_rcb->state = RCB_SENDUP;
192 
193 	mp->b_wptr = mp->b_rptr + pkt_len;
194 	mp->b_next = mp->b_cont = NULL;
195 
196 	/*
197 	 * Strip off one free rx control block from the free list
198 	 */
199 	free_index = rx_ring->rcb_head;
200 	free_rcb = rx_ring->free_list[free_index];
201 	ASSERT(free_rcb != NULL);
202 	rx_ring->free_list[free_index] = NULL;
203 	rx_ring->rcb_head = NEXT_INDEX(free_index, 1, rx_ring->free_list_size);
204 
205 	/*
206 	 * Put the rx control block to the work list
207 	 */
208 	rx_ring->work_list[index] = free_rcb;
209 
210 	return (mp);
211 }
212 
213 /*
214  * igb_rx_assoc_hcksum
215  *
216  * Check the rx hardware checksum status and associate the hcksum flags
217  */
218 static void
219 igb_rx_assoc_hcksum(mblk_t *mp, uint32_t status_error)
220 {
221 	uint32_t hcksum_flags = 0;
222 
223 	/* Ignore Checksum Indication */
224 	if (status_error & E1000_RXD_STAT_IXSM)
225 		return;
226 
227 	/*
228 	 * Check TCP/UDP checksum
229 	 */
230 	if (((status_error & E1000_RXD_STAT_TCPCS) ||
231 	    (status_error & E1000_RXD_STAT_UDPCS)) &&
232 	    !(status_error & E1000_RXDEXT_STATERR_TCPE))
233 		hcksum_flags |= HCK_FULLCKSUM | HCK_FULLCKSUM_OK;
234 
235 	/*
236 	 * Check IP Checksum
237 	 */
238 	if ((status_error & E1000_RXD_STAT_IPCS) &&
239 	    !(status_error & E1000_RXDEXT_STATERR_IPE))
240 		hcksum_flags |= HCK_IPV4_HDRCKSUM;
241 
242 	if (hcksum_flags != 0) {
243 		(void) hcksum_assoc(mp,
244 		    NULL, NULL, 0, 0, 0, 0, hcksum_flags, 0);
245 	}
246 }
247 
248 mblk_t *
249 igb_rx_ring_poll(void *arg, int bytes)
250 {
251 	igb_rx_ring_t *rx_ring = (igb_rx_ring_t *)arg;
252 	mblk_t *mp = NULL;
253 
254 	ASSERT(bytes >= 0);
255 
256 	if (bytes == 0)
257 		return (mp);
258 
259 	mutex_enter(&rx_ring->rx_lock);
260 	mp = igb_rx(rx_ring, bytes);
261 	mutex_exit(&rx_ring->rx_lock);
262 
263 	return (mp);
264 }
265 
266 /*
267  * igb_rx - Receive the data of one ring
268  *
269  * This function goes throught h/w descriptor in one specified rx ring,
270  * receives the data if the descriptor status shows the data is ready.
271  * It returns a chain of mblks containing the received data, to be
272  * passed up to mac_rx().
273  */
274 mblk_t *
275 igb_rx(igb_rx_ring_t *rx_ring, int poll_bytes)
276 {
277 	union e1000_adv_rx_desc *current_rbd;
278 	rx_control_block_t *current_rcb;
279 	mblk_t *mp;
280 	mblk_t *mblk_head;
281 	mblk_t **mblk_tail;
282 	uint32_t rx_next;
283 	uint32_t rx_tail;
284 	uint32_t pkt_len;
285 	uint32_t status_error;
286 	uint32_t pkt_num;
287 	uint32_t total_bytes;
288 	igb_t *igb = rx_ring->igb;
289 
290 	mblk_head = NULL;
291 	mblk_tail = &mblk_head;
292 
293 	/*
294 	 * Sync the receive descriptors before
295 	 * accepting the packets
296 	 */
297 	DMA_SYNC(&rx_ring->rbd_area, DDI_DMA_SYNC_FORKERNEL);
298 
299 	if (igb_check_dma_handle(
300 	    rx_ring->rbd_area.dma_handle) != DDI_FM_OK) {
301 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
302 	}
303 
304 	/*
305 	 * Get the start point of rx bd ring which should be examined
306 	 * during this cycle.
307 	 */
308 	rx_next = rx_ring->rbd_next;
309 
310 	current_rbd = &rx_ring->rbd_ring[rx_next];
311 	pkt_num = 0;
312 	total_bytes = 0;
313 	status_error = current_rbd->wb.upper.status_error;
314 	while (status_error & E1000_RXD_STAT_DD) {
315 		/*
316 		 * If hardware has found the errors, but the error
317 		 * is hardware checksum error, here does not discard the
318 		 * packet, and let upper layer compute the checksum;
319 		 * Otherwise discard the packet.
320 		 */
321 		if ((status_error & E1000_RXDEXT_ERR_FRAME_ERR_MASK) ||
322 		    !(status_error & E1000_RXD_STAT_EOP)) {
323 			IGB_DEBUG_STAT(rx_ring->stat_frame_error);
324 			goto rx_discard;
325 		}
326 
327 		IGB_DEBUG_STAT_COND(rx_ring->stat_cksum_error,
328 		    (status_error & E1000_RXDEXT_STATERR_TCPE) ||
329 		    (status_error & E1000_RXDEXT_STATERR_IPE));
330 
331 		pkt_len = current_rbd->wb.upper.length;
332 
333 		if ((poll_bytes != IGB_NO_POLL) &&
334 		    ((pkt_len + total_bytes) > poll_bytes))
335 			break;
336 
337 		IGB_DEBUG_STAT(rx_ring->stat_pkt_cnt);
338 		total_bytes += pkt_len;
339 
340 		mp = NULL;
341 		/*
342 		 * For packets with length more than the copy threshold,
343 		 * we'll firstly try to use the existed DMA buffer to built
344 		 * a mblk and send the mblk upstream.
345 		 *
346 		 * If the first method fails, or the packet length is less
347 		 * than the copy threshold, we'll allocate a new mblk and
348 		 * copy the packet data to the mblk.
349 		 */
350 		if (pkt_len > rx_ring->copy_thresh)
351 			mp = igb_rx_bind(rx_ring, rx_next, pkt_len);
352 
353 		if (mp == NULL)
354 			mp = igb_rx_copy(rx_ring, rx_next, pkt_len);
355 
356 		if (mp != NULL) {
357 			/*
358 			 * Check h/w checksum offload status
359 			 */
360 			if (igb->rx_hcksum_enable)
361 				igb_rx_assoc_hcksum(mp, status_error);
362 
363 			*mblk_tail = mp;
364 			mblk_tail = &mp->b_next;
365 		}
366 
367 rx_discard:
368 		/*
369 		 * Reset rx descriptor read bits
370 		 */
371 		current_rcb = rx_ring->work_list[rx_next];
372 		current_rbd->read.pkt_addr = current_rcb->rx_buf.dma_address;
373 		current_rbd->read.hdr_addr = 0;
374 
375 		rx_next = NEXT_INDEX(rx_next, 1, rx_ring->ring_size);
376 
377 		/*
378 		 * The receive function is in interrupt context, so here
379 		 * limit_per_intr is used to avoid doing receiving too long
380 		 * per interrupt.
381 		 */
382 		if (++pkt_num > rx_ring->limit_per_intr) {
383 			IGB_DEBUG_STAT(rx_ring->stat_exceed_pkt);
384 			break;
385 		}
386 
387 		current_rbd = &rx_ring->rbd_ring[rx_next];
388 		status_error = current_rbd->wb.upper.status_error;
389 	}
390 
391 	DMA_SYNC(&rx_ring->rbd_area, DDI_DMA_SYNC_FORDEV);
392 
393 	rx_ring->rbd_next = rx_next;
394 
395 	/*
396 	 * Update the h/w tail accordingly
397 	 */
398 	rx_tail = PREV_INDEX(rx_next, 1, rx_ring->ring_size);
399 
400 	E1000_WRITE_REG(&igb->hw, E1000_RDT(rx_ring->index), rx_tail);
401 
402 	if (igb_check_acc_handle(igb->osdep.reg_handle) != DDI_FM_OK) {
403 		ddi_fm_service_impact(igb->dip, DDI_SERVICE_DEGRADED);
404 	}
405 
406 	return (mblk_head);
407 }
408