xref: /titanic_41/usr/src/uts/common/io/bge/bge_recv2.c (revision 2a8d6eba033e4713ab12b61178f0513f1f075482)
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 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include "bge_impl.h"
28 
29 #define	U32TOPTR(x)	((void *)(uintptr_t)(uint32_t)(x))
30 #define	PTRTOU32(x)	((uint32_t)(uintptr_t)(void *)(x))
31 
32 /*
33  * ========== RX side routines ==========
34  */
35 
36 #define	BGE_DBG		BGE_DBG_RECV	/* debug flag for this code	*/
37 
38 static void bge_refill(bge_t *bgep, buff_ring_t *brp, sw_rbd_t *srbdp);
39 #pragma	inline(bge_refill)
40 
41 /*
42  * Return the specified buffer (srbdp) to the ring it came from (brp).
43  *
44  * Note:
45  *	If the driver is compiled with only one buffer ring *and* one
46  *	return ring, then the buffers must be returned in sequence.
47  *	In this case, we don't have to consider anything about the
48  *	buffer at all; we can simply advance the cyclic counter.  And
49  *	we don't even need the refill mutex <rf_lock>, as the caller
50  *	will already be holding the (one-and-only) <rx_lock>.
51  *
52  *	If the driver supports multiple buffer rings, but only one
53  *	return ring, the same still applies (to each buffer ring
54  *	separately).
55  */
56 static void
57 bge_refill(bge_t *bgep, buff_ring_t *brp, sw_rbd_t *srbdp)
58 {
59 	uint64_t slot;
60 
61 	_NOTE(ARGUNUSED(srbdp))
62 
63 	slot = brp->rf_next;
64 	brp->rf_next = NEXT(slot, brp->desc.nslots);
65 	bge_mbx_put(bgep, brp->chip_mbx_reg, slot);
66 }
67 
68 static mblk_t *bge_receive_packet(bge_t *bgep, bge_rbd_t *hw_rbd_p);
69 #pragma	inline(bge_receive_packet)
70 
71 static mblk_t *
72 bge_receive_packet(bge_t *bgep, bge_rbd_t *hw_rbd_p)
73 {
74 	bge_rbd_t hw_rbd;
75 	buff_ring_t *brp;
76 	sw_rbd_t *srbdp;
77 	uchar_t *dp;
78 	mblk_t *mp;
79 	uint_t len;
80 	uint_t minsize;
81 	uint_t maxsize;
82 	uint32_t pflags;
83 
84 	mp = NULL;
85 	hw_rbd = *hw_rbd_p;
86 
87 	switch (hw_rbd.flags & (RBD_FLAG_MINI_RING|RBD_FLAG_JUMBO_RING)) {
88 	case RBD_FLAG_MINI_RING|RBD_FLAG_JUMBO_RING:
89 	default:
90 		/* error, this shouldn't happen */
91 		BGE_PKTDUMP((bgep, &hw_rbd, NULL, "bad ring flags!"));
92 		goto error;
93 
94 	case RBD_FLAG_JUMBO_RING:
95 		brp = &bgep->buff[BGE_JUMBO_BUFF_RING];
96 		break;
97 
98 #if	(BGE_BUFF_RINGS_USED > 2)
99 	case RBD_FLAG_MINI_RING:
100 		brp = &bgep->buff[BGE_MINI_BUFF_RING];
101 		break;
102 #endif	/* BGE_BUFF_RINGS_USED > 2 */
103 
104 	case 0:
105 		brp = &bgep->buff[BGE_STD_BUFF_RING];
106 		break;
107 	}
108 
109 	if (hw_rbd.index >= brp->desc.nslots) {
110 		/* error, this shouldn't happen */
111 		BGE_PKTDUMP((bgep, &hw_rbd, NULL, "bad ring index!"));
112 		goto error;
113 	}
114 
115 	srbdp = &brp->sw_rbds[hw_rbd.index];
116 	if (hw_rbd.opaque != srbdp->pbuf.token) {
117 		/* bogus, drop the packet */
118 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "bad ring token"));
119 		goto refill;
120 	}
121 
122 	if ((hw_rbd.flags & RBD_FLAG_PACKET_END) == 0) {
123 		/* bogus, drop the packet */
124 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "unterminated packet"));
125 		goto refill;
126 	}
127 
128 	if (hw_rbd.flags & RBD_FLAG_FRAME_HAS_ERROR) {
129 		/* bogus, drop the packet */
130 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "errored packet"));
131 		goto refill;
132 	}
133 
134 	len = hw_rbd.len;
135 
136 #ifdef BGE_IPMI_ASF
137 	/*
138 	 * When IPMI/ASF is enabled, VLAN tag must be stripped.
139 	 */
140 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG))
141 		maxsize = bgep->chipid.ethmax_size + ETHERFCSL;
142 	else
143 #endif
144 		/*
145 		 * H/W will not strip the VLAN tag from incoming packet
146 		 * now, as RECEIVE_MODE_KEEP_VLAN_TAG bit is set in
147 		 * RECEIVE_MAC_MODE_REG register.
148 		 */
149 		maxsize = bgep->chipid.ethmax_size + VLAN_TAGSZ + ETHERFCSL;
150 	if (len > maxsize) {
151 		/* bogus, drop the packet */
152 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "oversize packet"));
153 		goto refill;
154 	}
155 
156 #ifdef BGE_IPMI_ASF
157 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG))
158 		minsize = ETHERMIN + ETHERFCSL - VLAN_TAGSZ;
159 	else
160 #endif
161 		minsize = ETHERMIN + ETHERFCSL;
162 	if (len < minsize) {
163 		/* bogus, drop the packet */
164 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "undersize packet"));
165 		goto refill;
166 	}
167 
168 	/*
169 	 * Packet looks good; get a buffer to copy it into.
170 	 * We want to leave some space at the front of the allocated
171 	 * buffer in case any upstream modules want to prepend some
172 	 * sort of header.  This also has the side-effect of making
173 	 * the packet *contents* 4-byte aligned, as required by NCA!
174 	 */
175 #ifdef BGE_IPMI_ASF
176 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
177 		mp = allocb(BGE_HEADROOM + len + VLAN_TAGSZ, 0);
178 	} else {
179 #endif
180 
181 		mp = allocb(BGE_HEADROOM + len, 0);
182 #ifdef BGE_IPMI_ASF
183 	}
184 #endif
185 	if (mp == NULL) {
186 		/* Nothing to do but drop the packet */
187 		goto refill;
188 	}
189 
190 	/*
191 	 * Sync the data and copy it to the STREAMS buffer.
192 	 */
193 	DMA_SYNC(srbdp->pbuf, DDI_DMA_SYNC_FORKERNEL);
194 	if (bge_check_dma_handle(bgep, srbdp->pbuf.dma_hdl) != DDI_FM_OK) {
195 		bgep->bge_dma_error = B_TRUE;
196 		bgep->bge_chip_state = BGE_CHIP_ERROR;
197 		return (NULL);
198 	}
199 #ifdef BGE_IPMI_ASF
200 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
201 		/*
202 		 * As VLAN tag has been stripped from incoming packet in ASF
203 		 * scenario, we insert it into this packet again.
204 		 */
205 		struct ether_vlan_header *ehp;
206 		mp->b_rptr = dp = mp->b_rptr + BGE_HEADROOM - VLAN_TAGSZ;
207 		bcopy(DMA_VPTR(srbdp->pbuf), dp, 2 * ETHERADDRL);
208 		ehp = (void *)dp;
209 		ehp->ether_tpid = ntohs(ETHERTYPE_VLAN);
210 		ehp->ether_tci = ntohs(hw_rbd.vlan_tci);
211 		bcopy(((uchar_t *)(DMA_VPTR(srbdp->pbuf))) + 2 * ETHERADDRL,
212 		    dp + 2 * ETHERADDRL + VLAN_TAGSZ,
213 		    len - 2 * ETHERADDRL);
214 	} else {
215 #endif
216 		mp->b_rptr = dp = mp->b_rptr + BGE_HEADROOM;
217 		bcopy(DMA_VPTR(srbdp->pbuf), dp, len);
218 #ifdef BGE_IPMI_ASF
219 	}
220 
221 	if (bgep->asf_enabled && (hw_rbd.flags & RBD_FLAG_VLAN_TAG)) {
222 		mp->b_wptr = dp + len + VLAN_TAGSZ - ETHERFCSL;
223 	} else
224 #endif
225 		mp->b_wptr = dp + len - ETHERFCSL;
226 
227 	/*
228 	 * Special check for one specific type of data corruption;
229 	 * in a good packet, the first 8 bytes are *very* unlikely
230 	 * to be the same as the second 8 bytes ... but we let the
231 	 * packet through just in case.
232 	 */
233 	if (bcmp(dp, dp+8, 8) == 0)
234 		BGE_PKTDUMP((bgep, &hw_rbd, srbdp, "stuttered packet?"));
235 
236 	pflags = 0;
237 	if (hw_rbd.flags & RBD_FLAG_TCP_UDP_CHECKSUM)
238 		pflags |= HCK_FULLCKSUM;
239 	if (hw_rbd.flags & RBD_FLAG_IP_CHECKSUM)
240 		pflags |= HCK_IPV4_HDRCKSUM;
241 	if (pflags != 0)
242 		(void) hcksum_assoc(mp, NULL, NULL, 0, 0, 0,
243 		    hw_rbd.tcp_udp_cksum, pflags, 0);
244 
245 refill:
246 	/*
247 	 * Replace the buffer in the ring it came from ...
248 	 */
249 	bge_refill(bgep, brp, srbdp);
250 	return (mp);
251 
252 error:
253 	/*
254 	 * We come here if the integrity of the ring descriptors
255 	 * (rather than merely packet data) appears corrupted.
256 	 * The factotum will attempt to reset-and-recover.
257 	 */
258 	bgep->bge_chip_state = BGE_CHIP_ERROR;
259 	bge_fm_ereport(bgep, DDI_FM_DEVICE_INVAL_STATE);
260 	return (NULL);
261 }
262 
263 /*
264  * Accept the packets received in the specified ring up to
265  * (but not including) the producer index in the status block.
266  *
267  * Returns a chain of mblks containing the received data, to be
268  * passed up to gld_recv() (we can't call gld_recv() from here,
269  * 'cos we're holding the per-ring receive lock at this point).
270  *
271  * This function must advance (rrp->rx_next) and write it back to
272  * the chip to indicate the packets it has accepted from the ring.
273  */
274 static mblk_t *bge_receive_ring(bge_t *bgep, recv_ring_t *rrp);
275 #ifndef	DEBUG
276 #pragma	inline(bge_receive_ring)
277 #endif
278 
279 static mblk_t *
280 bge_receive_ring(bge_t *bgep, recv_ring_t *rrp)
281 {
282 	bge_rbd_t *hw_rbd_p;
283 	uint64_t slot;
284 	mblk_t *head;
285 	mblk_t **tail;
286 	mblk_t *mp;
287 	int recv_cnt = 0;
288 
289 	ASSERT(mutex_owned(rrp->rx_lock));
290 
291 	/*
292 	 * Sync (all) the receive ring descriptors
293 	 * before accepting the packets they describe
294 	 */
295 	DMA_SYNC(rrp->desc, DDI_DMA_SYNC_FORKERNEL);
296 	if (*rrp->prod_index_p >= rrp->desc.nslots) {
297 		bgep->bge_chip_state = BGE_CHIP_ERROR;
298 		bge_fm_ereport(bgep, DDI_FM_DEVICE_INVAL_STATE);
299 		return (NULL);
300 	}
301 	if (bge_check_dma_handle(bgep, rrp->desc.dma_hdl) != DDI_FM_OK) {
302 		rrp->rx_next = *rrp->prod_index_p;
303 		bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
304 		bgep->bge_dma_error = B_TRUE;
305 		bgep->bge_chip_state = BGE_CHIP_ERROR;
306 		return (NULL);
307 	}
308 
309 	hw_rbd_p = DMA_VPTR(rrp->desc);
310 	head = NULL;
311 	tail = &head;
312 	slot = rrp->rx_next;
313 
314 	while ((slot != *rrp->prod_index_p) && /* Note: volatile	*/
315 	    (recv_cnt < BGE_MAXPKT_RCVED)) {
316 		if ((mp = bge_receive_packet(bgep, &hw_rbd_p[slot])) != NULL) {
317 			*tail = mp;
318 			tail = &mp->b_next;
319 			recv_cnt++;
320 		}
321 		rrp->rx_next = slot = NEXT(slot, rrp->desc.nslots);
322 	}
323 
324 	bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
325 	if (bge_check_acc_handle(bgep, bgep->io_handle) != DDI_FM_OK)
326 		bgep->bge_chip_state = BGE_CHIP_ERROR;
327 	return (head);
328 }
329 
330 /*
331  * XXX: Poll a particular ring. The implementation is incomplete.
332  * Once the ring interrupts are disabled, we need to do bge_recyle()
333  * for the ring as well and re enable the ring interrupt automatically
334  * if the poll doesn't find any packets in the ring. We need to
335  * have MSI-X interrupts support for this.
336  *
337  * The basic poll policy is that rings that are dealing with explicit
338  * flows (like TCP or some service) and are marked as such should
339  * have their own MSI-X interrupt per ring. bge_intr() should leave
340  * that interrupt disabled after an upcall. The ring is in poll mode.
341  * When a poll thread comes down and finds nothing, the MSI-X interrupt
342  * is automatically enabled. Squeue needs to deal with the race of
343  * a new interrupt firing and reaching before poll thread returns.
344  */
345 mblk_t *
346 bge_poll_ring(void *arg, int bytes_to_pickup)
347 {
348 	recv_ring_t *rrp = arg;
349 	bge_t *bgep = rrp->bgep;
350 	bge_rbd_t *hw_rbd_p;
351 	uint64_t slot;
352 	mblk_t *head;
353 	mblk_t **tail;
354 	mblk_t *mp;
355 	size_t sz = 0;
356 
357 	mutex_enter(rrp->rx_lock);
358 
359 	/*
360 	 * Sync (all) the receive ring descriptors
361 	 * before accepting the packets they describe
362 	 */
363 	DMA_SYNC(rrp->desc, DDI_DMA_SYNC_FORKERNEL);
364 	hw_rbd_p = DMA_VPTR(rrp->desc);
365 	head = NULL;
366 	tail = &head;
367 	slot = rrp->rx_next;
368 
369 	/* Note: volatile */
370 	while ((slot != *rrp->prod_index_p) && (sz <= bytes_to_pickup)) {
371 		if ((mp = bge_receive_packet(bgep, &hw_rbd_p[slot])) != NULL) {
372 			*tail = mp;
373 			sz += msgdsize(mp);
374 			tail = &mp->b_next;
375 		}
376 		rrp->rx_next = slot = NEXT(slot, rrp->desc.nslots);
377 	}
378 
379 	bge_mbx_put(bgep, rrp->chip_mbx_reg, rrp->rx_next);
380 	mutex_exit(rrp->rx_lock);
381 	return (head);
382 }
383 
384 /*
385  * Receive all packets in all rings.
386  */
387 void bge_receive(bge_t *bgep, bge_status_t *bsp);
388 #pragma	no_inline(bge_receive)
389 
390 void
391 bge_receive(bge_t *bgep, bge_status_t *bsp)
392 {
393 	recv_ring_t *rrp;
394 	uint64_t index;
395 	mblk_t *mp;
396 
397 	for (index = 0; index < bgep->chipid.rx_rings; index++) {
398 		/*
399 		 * Start from the first ring.
400 		 */
401 		rrp = &bgep->recv[index];
402 
403 		/*
404 		 * For each ring, (rrp->prod_index_p) points to the
405 		 * proper index within the status block (which has
406 		 * already been sync'd by the caller)
407 		 */
408 		ASSERT(rrp->prod_index_p == RECV_INDEX_P(bsp, index));
409 
410 		if (*rrp->prod_index_p == rrp->rx_next || rrp->poll_flag)
411 			continue;		/* no packets		*/
412 		if (mutex_tryenter(rrp->rx_lock) == 0)
413 			continue;		/* already in process	*/
414 		mp = bge_receive_ring(bgep, rrp);
415 		mutex_exit(rrp->rx_lock);
416 
417 		if (mp != NULL)
418 			mac_rx_ring(bgep->mh, rrp->ring_handle, mp,
419 			    rrp->ring_gen_num);
420 	}
421 }
422