xref: /openbmc/linux/drivers/infiniband/hw/hfi1/ud.c (revision 0984d159)
1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 
48 #include <linux/net.h>
49 #include <rdma/ib_smi.h>
50 
51 #include "hfi.h"
52 #include "mad.h"
53 #include "verbs_txreq.h"
54 #include "qp.h"
55 
56 /**
57  * ud_loopback - handle send on loopback QPs
58  * @sqp: the sending QP
59  * @swqe: the send work request
60  *
61  * This is called from hfi1_make_ud_req() to forward a WQE addressed
62  * to the same HFI.
63  * Note that the receive interrupt handler may be calling hfi1_ud_rcv()
64  * while this is being called.
65  */
66 static void ud_loopback(struct rvt_qp *sqp, struct rvt_swqe *swqe)
67 {
68 	struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
69 	struct hfi1_pportdata *ppd;
70 	struct rvt_qp *qp;
71 	struct ib_ah_attr *ah_attr;
72 	unsigned long flags;
73 	struct rvt_sge_state ssge;
74 	struct rvt_sge *sge;
75 	struct ib_wc wc;
76 	u32 length;
77 	enum ib_qp_type sqptype, dqptype;
78 
79 	rcu_read_lock();
80 
81 	qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
82 			    swqe->ud_wr.remote_qpn);
83 	if (!qp) {
84 		ibp->rvp.n_pkt_drops++;
85 		rcu_read_unlock();
86 		return;
87 	}
88 
89 	sqptype = sqp->ibqp.qp_type == IB_QPT_GSI ?
90 			IB_QPT_UD : sqp->ibqp.qp_type;
91 	dqptype = qp->ibqp.qp_type == IB_QPT_GSI ?
92 			IB_QPT_UD : qp->ibqp.qp_type;
93 
94 	if (dqptype != sqptype ||
95 	    !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
96 		ibp->rvp.n_pkt_drops++;
97 		goto drop;
98 	}
99 
100 	ah_attr = &ibah_to_rvtah(swqe->ud_wr.ah)->attr;
101 	ppd = ppd_from_ibp(ibp);
102 
103 	if (qp->ibqp.qp_num > 1) {
104 		u16 pkey;
105 		u16 slid;
106 		u8 sc5 = ibp->sl_to_sc[ah_attr->sl];
107 
108 		pkey = hfi1_get_pkey(ibp, sqp->s_pkey_index);
109 		slid = ppd->lid | (ah_attr->src_path_bits &
110 				   ((1 << ppd->lmc) - 1));
111 		if (unlikely(ingress_pkey_check(ppd, pkey, sc5,
112 						qp->s_pkey_index, slid))) {
113 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY, pkey,
114 				       ah_attr->sl,
115 				       sqp->ibqp.qp_num, qp->ibqp.qp_num,
116 				       slid, ah_attr->dlid);
117 			goto drop;
118 		}
119 	}
120 
121 	/*
122 	 * Check that the qkey matches (except for QP0, see 9.6.1.4.1).
123 	 * Qkeys with the high order bit set mean use the
124 	 * qkey from the QP context instead of the WR (see 10.2.5).
125 	 */
126 	if (qp->ibqp.qp_num) {
127 		u32 qkey;
128 
129 		qkey = (int)swqe->ud_wr.remote_qkey < 0 ?
130 			sqp->qkey : swqe->ud_wr.remote_qkey;
131 		if (unlikely(qkey != qp->qkey)) {
132 			u16 lid;
133 
134 			lid = ppd->lid | (ah_attr->src_path_bits &
135 					  ((1 << ppd->lmc) - 1));
136 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey,
137 				       ah_attr->sl,
138 				       sqp->ibqp.qp_num, qp->ibqp.qp_num,
139 				       lid,
140 				       ah_attr->dlid);
141 			goto drop;
142 		}
143 	}
144 
145 	/*
146 	 * A GRH is expected to precede the data even if not
147 	 * present on the wire.
148 	 */
149 	length = swqe->length;
150 	memset(&wc, 0, sizeof(wc));
151 	wc.byte_len = length + sizeof(struct ib_grh);
152 
153 	if (swqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
154 		wc.wc_flags = IB_WC_WITH_IMM;
155 		wc.ex.imm_data = swqe->wr.ex.imm_data;
156 	}
157 
158 	spin_lock_irqsave(&qp->r_lock, flags);
159 
160 	/*
161 	 * Get the next work request entry to find where to put the data.
162 	 */
163 	if (qp->r_flags & RVT_R_REUSE_SGE) {
164 		qp->r_flags &= ~RVT_R_REUSE_SGE;
165 	} else {
166 		int ret;
167 
168 		ret = hfi1_rvt_get_rwqe(qp, 0);
169 		if (ret < 0) {
170 			hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
171 			goto bail_unlock;
172 		}
173 		if (!ret) {
174 			if (qp->ibqp.qp_num == 0)
175 				ibp->rvp.n_vl15_dropped++;
176 			goto bail_unlock;
177 		}
178 	}
179 	/* Silently drop packets which are too big. */
180 	if (unlikely(wc.byte_len > qp->r_len)) {
181 		qp->r_flags |= RVT_R_REUSE_SGE;
182 		ibp->rvp.n_pkt_drops++;
183 		goto bail_unlock;
184 	}
185 
186 	if (ah_attr->ah_flags & IB_AH_GRH) {
187 		hfi1_copy_sge(&qp->r_sge, &ah_attr->grh,
188 			      sizeof(struct ib_grh), 1, 0);
189 		wc.wc_flags |= IB_WC_GRH;
190 	} else {
191 		hfi1_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
192 	}
193 	ssge.sg_list = swqe->sg_list + 1;
194 	ssge.sge = *swqe->sg_list;
195 	ssge.num_sge = swqe->wr.num_sge;
196 	sge = &ssge.sge;
197 	while (length) {
198 		u32 len = sge->length;
199 
200 		if (len > length)
201 			len = length;
202 		if (len > sge->sge_length)
203 			len = sge->sge_length;
204 		WARN_ON_ONCE(len == 0);
205 		hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, 1, 0);
206 		sge->vaddr += len;
207 		sge->length -= len;
208 		sge->sge_length -= len;
209 		if (sge->sge_length == 0) {
210 			if (--ssge.num_sge)
211 				*sge = *ssge.sg_list++;
212 		} else if (sge->length == 0 && sge->mr->lkey) {
213 			if (++sge->n >= RVT_SEGSZ) {
214 				if (++sge->m >= sge->mr->mapsz)
215 					break;
216 				sge->n = 0;
217 			}
218 			sge->vaddr =
219 				sge->mr->map[sge->m]->segs[sge->n].vaddr;
220 			sge->length =
221 				sge->mr->map[sge->m]->segs[sge->n].length;
222 		}
223 		length -= len;
224 	}
225 	rvt_put_ss(&qp->r_sge);
226 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
227 		goto bail_unlock;
228 	wc.wr_id = qp->r_wr_id;
229 	wc.status = IB_WC_SUCCESS;
230 	wc.opcode = IB_WC_RECV;
231 	wc.qp = &qp->ibqp;
232 	wc.src_qp = sqp->ibqp.qp_num;
233 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI) {
234 		if (sqp->ibqp.qp_type == IB_QPT_GSI ||
235 		    sqp->ibqp.qp_type == IB_QPT_SMI)
236 			wc.pkey_index = swqe->ud_wr.pkey_index;
237 		else
238 			wc.pkey_index = sqp->s_pkey_index;
239 	} else {
240 		wc.pkey_index = 0;
241 	}
242 	wc.slid = ppd->lid | (ah_attr->src_path_bits & ((1 << ppd->lmc) - 1));
243 	/* Check for loopback when the port lid is not set */
244 	if (wc.slid == 0 && sqp->ibqp.qp_type == IB_QPT_GSI)
245 		wc.slid = be16_to_cpu(IB_LID_PERMISSIVE);
246 	wc.sl = ah_attr->sl;
247 	wc.dlid_path_bits = ah_attr->dlid & ((1 << ppd->lmc) - 1);
248 	wc.port_num = qp->port_num;
249 	/* Signal completion event if the solicited bit is set. */
250 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
251 		     swqe->wr.send_flags & IB_SEND_SOLICITED);
252 	ibp->rvp.n_loop_pkts++;
253 bail_unlock:
254 	spin_unlock_irqrestore(&qp->r_lock, flags);
255 drop:
256 	rcu_read_unlock();
257 }
258 
259 /**
260  * hfi1_make_ud_req - construct a UD request packet
261  * @qp: the QP
262  *
263  * Assume s_lock is held.
264  *
265  * Return 1 if constructed; otherwise, return 0.
266  */
267 int hfi1_make_ud_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
268 {
269 	struct hfi1_qp_priv *priv = qp->priv;
270 	struct hfi1_other_headers *ohdr;
271 	struct ib_ah_attr *ah_attr;
272 	struct hfi1_pportdata *ppd;
273 	struct hfi1_ibport *ibp;
274 	struct rvt_swqe *wqe;
275 	u32 nwords;
276 	u32 extra_bytes;
277 	u32 bth0;
278 	u16 lrh0;
279 	u16 lid;
280 	int next_cur;
281 	u8 sc5;
282 
283 	ps->s_txreq = get_txreq(ps->dev, qp);
284 	if (IS_ERR(ps->s_txreq))
285 		goto bail_no_tx;
286 
287 	if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
288 		if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
289 			goto bail;
290 		/* We are in the error state, flush the work request. */
291 		smp_read_barrier_depends(); /* see post_one_send */
292 		if (qp->s_last == ACCESS_ONCE(qp->s_head))
293 			goto bail;
294 		/* If DMAs are in progress, we can't flush immediately. */
295 		if (iowait_sdma_pending(&priv->s_iowait)) {
296 			qp->s_flags |= RVT_S_WAIT_DMA;
297 			goto bail;
298 		}
299 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
300 		hfi1_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
301 		goto done_free_tx;
302 	}
303 
304 	/* see post_one_send() */
305 	smp_read_barrier_depends();
306 	if (qp->s_cur == ACCESS_ONCE(qp->s_head))
307 		goto bail;
308 
309 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
310 	next_cur = qp->s_cur + 1;
311 	if (next_cur >= qp->s_size)
312 		next_cur = 0;
313 
314 	/* Construct the header. */
315 	ibp = to_iport(qp->ibqp.device, qp->port_num);
316 	ppd = ppd_from_ibp(ibp);
317 	ah_attr = &ibah_to_rvtah(wqe->ud_wr.ah)->attr;
318 	if (ah_attr->dlid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
319 	    ah_attr->dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
320 		lid = ah_attr->dlid & ~((1 << ppd->lmc) - 1);
321 		if (unlikely(!loopback &&
322 			     (lid == ppd->lid ||
323 			      (lid == be16_to_cpu(IB_LID_PERMISSIVE) &&
324 			      qp->ibqp.qp_type == IB_QPT_GSI)))) {
325 			unsigned long tflags = ps->flags;
326 			/*
327 			 * If DMAs are in progress, we can't generate
328 			 * a completion for the loopback packet since
329 			 * it would be out of order.
330 			 * Instead of waiting, we could queue a
331 			 * zero length descriptor so we get a callback.
332 			 */
333 			if (iowait_sdma_pending(&priv->s_iowait)) {
334 				qp->s_flags |= RVT_S_WAIT_DMA;
335 				goto bail;
336 			}
337 			qp->s_cur = next_cur;
338 			spin_unlock_irqrestore(&qp->s_lock, tflags);
339 			ud_loopback(qp, wqe);
340 			spin_lock_irqsave(&qp->s_lock, tflags);
341 			ps->flags = tflags;
342 			hfi1_send_complete(qp, wqe, IB_WC_SUCCESS);
343 			goto done_free_tx;
344 		}
345 	}
346 
347 	qp->s_cur = next_cur;
348 	extra_bytes = -wqe->length & 3;
349 	nwords = (wqe->length + extra_bytes) >> 2;
350 
351 	/* header size in 32-bit words LRH+BTH+DETH = (8+12+8)/4. */
352 	qp->s_hdrwords = 7;
353 	qp->s_cur_size = wqe->length;
354 	qp->s_cur_sge = &qp->s_sge;
355 	qp->s_srate = ah_attr->static_rate;
356 	qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
357 	qp->s_wqe = wqe;
358 	qp->s_sge.sge = wqe->sg_list[0];
359 	qp->s_sge.sg_list = wqe->sg_list + 1;
360 	qp->s_sge.num_sge = wqe->wr.num_sge;
361 	qp->s_sge.total_len = wqe->length;
362 
363 	if (ah_attr->ah_flags & IB_AH_GRH) {
364 		/* Header size in 32-bit words. */
365 		qp->s_hdrwords += hfi1_make_grh(ibp,
366 						&ps->s_txreq->phdr.hdr.u.l.grh,
367 						&ah_attr->grh,
368 						qp->s_hdrwords, nwords);
369 		lrh0 = HFI1_LRH_GRH;
370 		ohdr = &ps->s_txreq->phdr.hdr.u.l.oth;
371 		/*
372 		 * Don't worry about sending to locally attached multicast
373 		 * QPs.  It is unspecified by the spec. what happens.
374 		 */
375 	} else {
376 		/* Header size in 32-bit words. */
377 		lrh0 = HFI1_LRH_BTH;
378 		ohdr = &ps->s_txreq->phdr.hdr.u.oth;
379 	}
380 	if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
381 		qp->s_hdrwords++;
382 		ohdr->u.ud.imm_data = wqe->wr.ex.imm_data;
383 		bth0 = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE << 24;
384 	} else {
385 		bth0 = IB_OPCODE_UD_SEND_ONLY << 24;
386 	}
387 	sc5 = ibp->sl_to_sc[ah_attr->sl];
388 	lrh0 |= (ah_attr->sl & 0xf) << 4;
389 	if (qp->ibqp.qp_type == IB_QPT_SMI) {
390 		lrh0 |= 0xF000; /* Set VL (see ch. 13.5.3.1) */
391 		priv->s_sc = 0xf;
392 	} else {
393 		lrh0 |= (sc5 & 0xf) << 12;
394 		priv->s_sc = sc5;
395 	}
396 	priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
397 	ps->s_txreq->sde = priv->s_sde;
398 	priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
399 	ps->s_txreq->psc = priv->s_sendcontext;
400 	ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
401 	ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(ah_attr->dlid);
402 	ps->s_txreq->phdr.hdr.lrh[2] =
403 		cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
404 	if (ah_attr->dlid == be16_to_cpu(IB_LID_PERMISSIVE)) {
405 		ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE;
406 	} else {
407 		lid = ppd->lid;
408 		if (lid) {
409 			lid |= ah_attr->src_path_bits & ((1 << ppd->lmc) - 1);
410 			ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(lid);
411 		} else {
412 			ps->s_txreq->phdr.hdr.lrh[3] = IB_LID_PERMISSIVE;
413 		}
414 	}
415 	if (wqe->wr.send_flags & IB_SEND_SOLICITED)
416 		bth0 |= IB_BTH_SOLICITED;
417 	bth0 |= extra_bytes << 20;
418 	if (qp->ibqp.qp_type == IB_QPT_GSI || qp->ibqp.qp_type == IB_QPT_SMI)
419 		bth0 |= hfi1_get_pkey(ibp, wqe->ud_wr.pkey_index);
420 	else
421 		bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
422 	ohdr->bth[0] = cpu_to_be32(bth0);
423 	ohdr->bth[1] = cpu_to_be32(wqe->ud_wr.remote_qpn);
424 	ohdr->bth[2] = cpu_to_be32(mask_psn(wqe->psn));
425 	/*
426 	 * Qkeys with the high order bit set mean use the
427 	 * qkey from the QP context instead of the WR (see 10.2.5).
428 	 */
429 	ohdr->u.ud.deth[0] = cpu_to_be32((int)wqe->ud_wr.remote_qkey < 0 ?
430 					 qp->qkey : wqe->ud_wr.remote_qkey);
431 	ohdr->u.ud.deth[1] = cpu_to_be32(qp->ibqp.qp_num);
432 	/* disarm any ahg */
433 	priv->s_hdr->ahgcount = 0;
434 	priv->s_hdr->ahgidx = 0;
435 	priv->s_hdr->tx_flags = 0;
436 	priv->s_hdr->sde = NULL;
437 	/* pbc */
438 	ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
439 
440 	return 1;
441 
442 done_free_tx:
443 	hfi1_put_txreq(ps->s_txreq);
444 	ps->s_txreq = NULL;
445 	return 1;
446 
447 bail:
448 	hfi1_put_txreq(ps->s_txreq);
449 
450 bail_no_tx:
451 	ps->s_txreq = NULL;
452 	qp->s_flags &= ~RVT_S_BUSY;
453 	qp->s_hdrwords = 0;
454 	return 0;
455 }
456 
457 /*
458  * Hardware can't check this so we do it here.
459  *
460  * This is a slightly different algorithm than the standard pkey check.  It
461  * special cases the management keys and allows for 0x7fff and 0xffff to be in
462  * the table at the same time.
463  *
464  * @returns the index found or -1 if not found
465  */
466 int hfi1_lookup_pkey_idx(struct hfi1_ibport *ibp, u16 pkey)
467 {
468 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
469 	unsigned i;
470 
471 	if (pkey == FULL_MGMT_P_KEY || pkey == LIM_MGMT_P_KEY) {
472 		unsigned lim_idx = -1;
473 
474 		for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i) {
475 			/* here we look for an exact match */
476 			if (ppd->pkeys[i] == pkey)
477 				return i;
478 			if (ppd->pkeys[i] == LIM_MGMT_P_KEY)
479 				lim_idx = i;
480 		}
481 
482 		/* did not find 0xffff return 0x7fff idx if found */
483 		if (pkey == FULL_MGMT_P_KEY)
484 			return lim_idx;
485 
486 		/* no match...  */
487 		return -1;
488 	}
489 
490 	pkey &= 0x7fff; /* remove limited/full membership bit */
491 
492 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); ++i)
493 		if ((ppd->pkeys[i] & 0x7fff) == pkey)
494 			return i;
495 
496 	/*
497 	 * Should not get here, this means hardware failed to validate pkeys.
498 	 */
499 	return -1;
500 }
501 
502 void return_cnp(struct hfi1_ibport *ibp, struct rvt_qp *qp, u32 remote_qpn,
503 		u32 pkey, u32 slid, u32 dlid, u8 sc5,
504 		const struct ib_grh *old_grh)
505 {
506 	u64 pbc, pbc_flags = 0;
507 	u32 bth0, plen, vl, hwords = 5;
508 	u16 lrh0;
509 	u8 sl = ibp->sc_to_sl[sc5];
510 	struct hfi1_ib_header hdr;
511 	struct hfi1_other_headers *ohdr;
512 	struct pio_buf *pbuf;
513 	struct send_context *ctxt = qp_to_send_context(qp, sc5);
514 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
515 
516 	if (old_grh) {
517 		struct ib_grh *grh = &hdr.u.l.grh;
518 
519 		grh->version_tclass_flow = old_grh->version_tclass_flow;
520 		grh->paylen = cpu_to_be16((hwords - 2 + SIZE_OF_CRC) << 2);
521 		grh->hop_limit = 0xff;
522 		grh->sgid = old_grh->dgid;
523 		grh->dgid = old_grh->sgid;
524 		ohdr = &hdr.u.l.oth;
525 		lrh0 = HFI1_LRH_GRH;
526 		hwords += sizeof(struct ib_grh) / sizeof(u32);
527 	} else {
528 		ohdr = &hdr.u.oth;
529 		lrh0 = HFI1_LRH_BTH;
530 	}
531 
532 	lrh0 |= (sc5 & 0xf) << 12 | sl << 4;
533 
534 	bth0 = pkey | (IB_OPCODE_CNP << 24);
535 	ohdr->bth[0] = cpu_to_be32(bth0);
536 
537 	ohdr->bth[1] = cpu_to_be32(remote_qpn | (1 << HFI1_BECN_SHIFT));
538 	ohdr->bth[2] = 0; /* PSN 0 */
539 
540 	hdr.lrh[0] = cpu_to_be16(lrh0);
541 	hdr.lrh[1] = cpu_to_be16(dlid);
542 	hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
543 	hdr.lrh[3] = cpu_to_be16(slid);
544 
545 	plen = 2 /* PBC */ + hwords;
546 	pbc_flags |= (!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT;
547 	vl = sc_to_vlt(ppd->dd, sc5);
548 	pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
549 	if (ctxt) {
550 		pbuf = sc_buffer_alloc(ctxt, plen, NULL, NULL);
551 		if (pbuf)
552 			ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
553 						 &hdr, hwords);
554 	}
555 }
556 
557 /*
558  * opa_smp_check() - Do the regular pkey checking, and the additional
559  * checks for SMPs specified in OPAv1 rev 0.90, section 9.10.26
560  * ("SMA Packet Checks").
561  *
562  * Note that:
563  *   - Checks are done using the pkey directly from the packet's BTH,
564  *     and specifically _not_ the pkey that we attach to the completion,
565  *     which may be different.
566  *   - These checks are specifically for "non-local" SMPs (i.e., SMPs
567  *     which originated on another node). SMPs which are sent from, and
568  *     destined to this node are checked in opa_local_smp_check().
569  *
570  * At the point where opa_smp_check() is called, we know:
571  *   - destination QP is QP0
572  *
573  * opa_smp_check() returns 0 if all checks succeed, 1 otherwise.
574  */
575 static int opa_smp_check(struct hfi1_ibport *ibp, u16 pkey, u8 sc5,
576 			 struct rvt_qp *qp, u16 slid, struct opa_smp *smp)
577 {
578 	struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
579 
580 	/*
581 	 * I don't think it's possible for us to get here with sc != 0xf,
582 	 * but check it to be certain.
583 	 */
584 	if (sc5 != 0xf)
585 		return 1;
586 
587 	if (rcv_pkey_check(ppd, pkey, sc5, slid))
588 		return 1;
589 
590 	/*
591 	 * At this point we know (and so don't need to check again) that
592 	 * the pkey is either LIM_MGMT_P_KEY, or FULL_MGMT_P_KEY
593 	 * (see ingress_pkey_check).
594 	 */
595 	if (smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE &&
596 	    smp->mgmt_class != IB_MGMT_CLASS_SUBN_LID_ROUTED) {
597 		ingress_pkey_table_fail(ppd, pkey, slid);
598 		return 1;
599 	}
600 
601 	/*
602 	 * SMPs fall into one of four (disjoint) categories:
603 	 * SMA request, SMA response, trap, or trap repress.
604 	 * Our response depends, in part, on which type of
605 	 * SMP we're processing.
606 	 *
607 	 * If this is not an SMA request, or trap repress:
608 	 *   - accept MAD if the port is running an SM
609 	 *   - pkey == FULL_MGMT_P_KEY =>
610 	 *       reply with unsupported method (i.e., just mark
611 	 *       the smp's status field here, and let it be
612 	 *       processed normally)
613 	 *   - pkey != LIM_MGMT_P_KEY =>
614 	 *       increment port recv constraint errors, drop MAD
615 	 * If this is an SMA request or trap repress:
616 	 *   - pkey != FULL_MGMT_P_KEY =>
617 	 *       increment port recv constraint errors, drop MAD
618 	 */
619 	switch (smp->method) {
620 	case IB_MGMT_METHOD_GET:
621 	case IB_MGMT_METHOD_SET:
622 	case IB_MGMT_METHOD_REPORT:
623 	case IB_MGMT_METHOD_TRAP_REPRESS:
624 		if (pkey != FULL_MGMT_P_KEY) {
625 			ingress_pkey_table_fail(ppd, pkey, slid);
626 			return 1;
627 		}
628 		break;
629 	case IB_MGMT_METHOD_SEND:
630 	case IB_MGMT_METHOD_TRAP:
631 	case IB_MGMT_METHOD_GET_RESP:
632 	case IB_MGMT_METHOD_REPORT_RESP:
633 		if (ibp->rvp.port_cap_flags & IB_PORT_SM)
634 			return 0;
635 		if (pkey == FULL_MGMT_P_KEY) {
636 			smp->status |= IB_SMP_UNSUP_METHOD;
637 			return 0;
638 		}
639 		if (pkey != LIM_MGMT_P_KEY) {
640 			ingress_pkey_table_fail(ppd, pkey, slid);
641 			return 1;
642 		}
643 		break;
644 	default:
645 		break;
646 	}
647 	return 0;
648 }
649 
650 /**
651  * hfi1_ud_rcv - receive an incoming UD packet
652  * @ibp: the port the packet came in on
653  * @hdr: the packet header
654  * @rcv_flags: flags relevant to rcv processing
655  * @data: the packet data
656  * @tlen: the packet length
657  * @qp: the QP the packet came on
658  *
659  * This is called from qp_rcv() to process an incoming UD packet
660  * for the given QP.
661  * Called at interrupt level.
662  */
663 void hfi1_ud_rcv(struct hfi1_packet *packet)
664 {
665 	struct hfi1_other_headers *ohdr = packet->ohdr;
666 	int opcode;
667 	u32 hdrsize = packet->hlen;
668 	u32 pad;
669 	struct ib_wc wc;
670 	u32 qkey;
671 	u32 src_qp;
672 	u16 dlid, pkey;
673 	int mgmt_pkey_idx = -1;
674 	struct hfi1_ibport *ibp = &packet->rcd->ppd->ibport_data;
675 	struct hfi1_ib_header *hdr = packet->hdr;
676 	u32 rcv_flags = packet->rcv_flags;
677 	void *data = packet->ebuf;
678 	u32 tlen = packet->tlen;
679 	struct rvt_qp *qp = packet->qp;
680 	bool has_grh = rcv_flags & HFI1_HAS_GRH;
681 	u8 sc5 = hdr2sc((struct hfi1_message_header *)hdr, packet->rhf);
682 	u32 bth1;
683 	int is_mcast;
684 	struct ib_grh *grh = NULL;
685 
686 	qkey = be32_to_cpu(ohdr->u.ud.deth[0]);
687 	src_qp = be32_to_cpu(ohdr->u.ud.deth[1]) & RVT_QPN_MASK;
688 	dlid = be16_to_cpu(hdr->lrh[1]);
689 	is_mcast = (dlid > be16_to_cpu(IB_MULTICAST_LID_BASE)) &&
690 			(dlid != be16_to_cpu(IB_LID_PERMISSIVE));
691 	bth1 = be32_to_cpu(ohdr->bth[1]);
692 	if (unlikely(bth1 & HFI1_BECN_SMASK)) {
693 		/*
694 		 * In pre-B0 h/w the CNP_OPCODE is handled via an
695 		 * error path.
696 		 */
697 		struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
698 		u32 lqpn =  be32_to_cpu(ohdr->bth[1]) & RVT_QPN_MASK;
699 		u8 sl;
700 
701 		sl = ibp->sc_to_sl[sc5];
702 
703 		process_becn(ppd, sl, 0, lqpn, 0, IB_CC_SVCTYPE_UD);
704 	}
705 
706 	/*
707 	 * The opcode is in the low byte when its in network order
708 	 * (top byte when in host order).
709 	 */
710 	opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
711 	opcode &= 0xff;
712 
713 	pkey = (u16)be32_to_cpu(ohdr->bth[0]);
714 
715 	if (!is_mcast && (opcode != IB_OPCODE_CNP) && bth1 & HFI1_FECN_SMASK) {
716 		u16 slid = be16_to_cpu(hdr->lrh[3]);
717 
718 		return_cnp(ibp, qp, src_qp, pkey, dlid, slid, sc5, grh);
719 	}
720 	/*
721 	 * Get the number of bytes the message was padded by
722 	 * and drop incomplete packets.
723 	 */
724 	pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
725 	if (unlikely(tlen < (hdrsize + pad + 4)))
726 		goto drop;
727 
728 	tlen -= hdrsize + pad + 4;
729 
730 	/*
731 	 * Check that the permissive LID is only used on QP0
732 	 * and the QKEY matches (see 9.6.1.4.1 and 9.6.1.5.1).
733 	 */
734 	if (qp->ibqp.qp_num) {
735 		if (unlikely(hdr->lrh[1] == IB_LID_PERMISSIVE ||
736 			     hdr->lrh[3] == IB_LID_PERMISSIVE))
737 			goto drop;
738 		if (qp->ibqp.qp_num > 1) {
739 			struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
740 			u16 slid;
741 
742 			slid = be16_to_cpu(hdr->lrh[3]);
743 			if (unlikely(rcv_pkey_check(ppd, pkey, sc5, slid))) {
744 				/*
745 				 * Traps will not be sent for packets dropped
746 				 * by the HW. This is fine, as sending trap
747 				 * for invalid pkeys is optional according to
748 				 * IB spec (release 1.3, section 10.9.4)
749 				 */
750 				hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
751 					       pkey,
752 					       (be16_to_cpu(hdr->lrh[0]) >> 4) &
753 						0xF,
754 					       src_qp, qp->ibqp.qp_num,
755 					       be16_to_cpu(hdr->lrh[3]),
756 					       be16_to_cpu(hdr->lrh[1]));
757 				return;
758 			}
759 		} else {
760 			/* GSI packet */
761 			mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
762 			if (mgmt_pkey_idx < 0)
763 				goto drop;
764 		}
765 		if (unlikely(qkey != qp->qkey)) {
766 			hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_Q_KEY, qkey,
767 				       (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
768 				       src_qp, qp->ibqp.qp_num,
769 				       be16_to_cpu(hdr->lrh[3]),
770 				       be16_to_cpu(hdr->lrh[1]));
771 			return;
772 		}
773 		/* Drop invalid MAD packets (see 13.5.3.1). */
774 		if (unlikely(qp->ibqp.qp_num == 1 &&
775 			     (tlen > 2048 ||
776 			      (be16_to_cpu(hdr->lrh[0]) >> 12) == 15)))
777 			goto drop;
778 	} else {
779 		/* Received on QP0, and so by definition, this is an SMP */
780 		struct opa_smp *smp = (struct opa_smp *)data;
781 		u16 slid = be16_to_cpu(hdr->lrh[3]);
782 
783 		if (opa_smp_check(ibp, pkey, sc5, qp, slid, smp))
784 			goto drop;
785 
786 		if (tlen > 2048)
787 			goto drop;
788 		if ((hdr->lrh[1] == IB_LID_PERMISSIVE ||
789 		     hdr->lrh[3] == IB_LID_PERMISSIVE) &&
790 		    smp->mgmt_class != IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
791 			goto drop;
792 
793 		/* look up SMI pkey */
794 		mgmt_pkey_idx = hfi1_lookup_pkey_idx(ibp, pkey);
795 		if (mgmt_pkey_idx < 0)
796 			goto drop;
797 	}
798 
799 	if (qp->ibqp.qp_num > 1 &&
800 	    opcode == IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE) {
801 		wc.ex.imm_data = ohdr->u.ud.imm_data;
802 		wc.wc_flags = IB_WC_WITH_IMM;
803 		tlen -= sizeof(u32);
804 	} else if (opcode == IB_OPCODE_UD_SEND_ONLY) {
805 		wc.ex.imm_data = 0;
806 		wc.wc_flags = 0;
807 	} else {
808 		goto drop;
809 	}
810 
811 	/*
812 	 * A GRH is expected to precede the data even if not
813 	 * present on the wire.
814 	 */
815 	wc.byte_len = tlen + sizeof(struct ib_grh);
816 
817 	/*
818 	 * Get the next work request entry to find where to put the data.
819 	 */
820 	if (qp->r_flags & RVT_R_REUSE_SGE) {
821 		qp->r_flags &= ~RVT_R_REUSE_SGE;
822 	} else {
823 		int ret;
824 
825 		ret = hfi1_rvt_get_rwqe(qp, 0);
826 		if (ret < 0) {
827 			hfi1_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
828 			return;
829 		}
830 		if (!ret) {
831 			if (qp->ibqp.qp_num == 0)
832 				ibp->rvp.n_vl15_dropped++;
833 			return;
834 		}
835 	}
836 	/* Silently drop packets which are too big. */
837 	if (unlikely(wc.byte_len > qp->r_len)) {
838 		qp->r_flags |= RVT_R_REUSE_SGE;
839 		goto drop;
840 	}
841 	if (has_grh) {
842 		hfi1_copy_sge(&qp->r_sge, &hdr->u.l.grh,
843 			      sizeof(struct ib_grh), 1, 0);
844 		wc.wc_flags |= IB_WC_GRH;
845 	} else {
846 		hfi1_skip_sge(&qp->r_sge, sizeof(struct ib_grh), 1);
847 	}
848 	hfi1_copy_sge(&qp->r_sge, data, wc.byte_len - sizeof(struct ib_grh),
849 		      1, 0);
850 	rvt_put_ss(&qp->r_sge);
851 	if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
852 		return;
853 	wc.wr_id = qp->r_wr_id;
854 	wc.status = IB_WC_SUCCESS;
855 	wc.opcode = IB_WC_RECV;
856 	wc.vendor_err = 0;
857 	wc.qp = &qp->ibqp;
858 	wc.src_qp = src_qp;
859 
860 	if (qp->ibqp.qp_type == IB_QPT_GSI ||
861 	    qp->ibqp.qp_type == IB_QPT_SMI) {
862 		if (mgmt_pkey_idx < 0) {
863 			if (net_ratelimit()) {
864 				struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
865 				struct hfi1_devdata *dd = ppd->dd;
866 
867 				dd_dev_err(dd, "QP type %d mgmt_pkey_idx < 0 and packet not dropped???\n",
868 					   qp->ibqp.qp_type);
869 				mgmt_pkey_idx = 0;
870 			}
871 		}
872 		wc.pkey_index = (unsigned)mgmt_pkey_idx;
873 	} else {
874 		wc.pkey_index = 0;
875 	}
876 
877 	wc.slid = be16_to_cpu(hdr->lrh[3]);
878 	wc.sl = ibp->sc_to_sl[sc5];
879 
880 	/*
881 	 * Save the LMC lower bits if the destination LID is a unicast LID.
882 	 */
883 	wc.dlid_path_bits = dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE) ? 0 :
884 		dlid & ((1 << ppd_from_ibp(ibp)->lmc) - 1);
885 	wc.port_num = qp->port_num;
886 	/* Signal completion event if the solicited bit is set. */
887 	rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
888 		     (ohdr->bth[0] &
889 		      cpu_to_be32(IB_BTH_SOLICITED)) != 0);
890 	return;
891 
892 drop:
893 	ibp->rvp.n_pkt_drops++;
894 }
895